Datasets:

jamescalam

/

ai-arxiv-chunked

like 32

2209.07686

363

TÞÑAhockey player Kyle Palmieri is.Called forslash ingispartofhockey being. VAÞÑyes qPrompt 6 åQÞÑIsthefollowingsentence plausible? “Car sonWentz setthepick androll.” TÞÑAnAmer icanfootballplayer CarsonWentz is.Partofbasketballpick androllis,notfootball. VAÞÑno qPrompt 7 åQÞÑIsthefollowingsentence plausible? “Mal colm Brog donbanked theshot in.” TÞÑAbasketballplayer Malcolm Brog donis.Inispartofbasketballbank ingtheshot. VAÞÑyes qPrompt 8 åQÞÑIsthefollowingsentence plausible? “Dray mond Green threw atouch down.” TÞÑAnbasketballplayer Dray mond Green is.Atouch down ispartoffootballthrow ing,notbasketball. VAÞÑno 102 Table 72: D ATEwith thoughts in YodaSpeak. qPrompt 1

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

364

VAÞÑno 102 Table 72: D ATEwith thoughts in YodaSpeak. qPrompt 1 åQÞÑItis4/19/1969 today. What isthedate 24hours later inMM/D D/YYYY? TÞÑ04/19/1969, todayis.Later isonedayaftertoday, 24hours, 04/20/1969, which would be. VAÞÑ04/20/1969 qPrompt 2 åQÞÑTheconcertwassched uled tobeon06/01/1943, butwasdelayed byonedaytotoday.What isthe date 10days agoinMM/D D/YYYY? TÞÑ06/02/1943, onedayafter06/01/1943 is,06/02/1943, sotodayis.05/23/1943, 10days before todayis. VAÞÑ05/23/1943 qPrompt 3 åQÞÑThefirstdayof2019 isaTues day, andtodayisthefirstMon dayof2019. What isthedate today inMM/D D/YYYY?

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

365

inMM/D D/YYYY? TÞÑTues day, ifthefirstdayof2019 was, aTues day, then 01/01/2019 was. ThefirstMon day, today is,sixdays later, would be.01/07/2019, sotodayis. VAÞÑ01/07/2019 qPrompt 4 åQÞÑJane wasborn onthelastdayofFebru aryin2001. Todayisher16-year-oldbirth day. What isthe date yesterdayinMM/D D/YYYY? TÞÑThe28th, thelastdayofFebru aryis,Born on02/28/2001, soJane was. Her16-year oldbirth day, todayis,02/28/2017, sotodayis.02/27/2017, soyesterdaywas. VAÞÑ02/27/2017 qPrompt 5 åQÞÑ2015 iscomingin36hours. What isthedate oneweek from todayinMM/D D/YYYY? TÞÑCom ingin36hours, if2015 is,comingin2days, then itis.12/30/2014, 2days before01/01/2015

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

366

TÞÑCom ingin36hours, if2015 is,comingin2days, then itis.12/30/2014, 2days before01/01/2015 is,12/30/2014, sotodayis.01/05/2015, sooneweek from todaywillbe. VAÞÑ01/05/2015 qPrompt 6 åQÞÑJane thought todayis3/11/2002, buttodayisinfactMar 12,which is1daylater. What isthedate 24hours later inMM/D D/YYYY? TÞÑ03/12/2002, todayis.Later willbe03/13/2002, sothedate 24hours. VAÞÑ03/13/2002 103 Table 73: GSM-8 Kwith examples in YodaSpeak. qPrompt 1 åQÞÑFive toys, shawn has.ForChrist mas, twotoys each from hismom anddad, hegot.Hehave now, how many toys does? TÞÑWith 5toys, Shawn started. 2toys each from hismom anddad, ifhegot,then thatis4more toys. 5+4=9. VAÞÑ9 qPrompt 2

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

367

5+4=9. VAÞÑ9 qPrompt 2 åQÞÑ3carsintheparkinglotand2more carsarrive, ifthere are,In theparkinglot,how many carsare? TÞÑOriginally 3cars, there are.2more carsarrive. 3+2=5. VAÞÑ5 qPrompt 3 åQÞÑ20lollipops, Jasonhad. Denny some lollipops, hegave. 12lollipops, now Jasonhas.Jasongive toDenny, how many lollipops did? TÞÑWith 20lollipops, Jasonstarted. 12aftergivingsome toDenny, then hehad. Sohegave Denny 20-12=8. VAÞÑ8 qPrompt 4 åQÞÑNine computersintheserver room, there were. Installed each day, fivemore computerswere, from Mon daytoThurs day. Now intheserver room, how many computersare? TÞÑOriginally 9computers,there were. Foreach of4days, added, 5more computerswere. So5*4 =20computerswere added. 9+20is29.

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

368

=20computerswere added. 9+20is29. VAÞÑ29 qPrompt 5 åQÞÑ15trees inthegrove, there are.Grove work erswillplant trees inthegrove today.Done, afterthey are,21 trees, there willbe.Thegrove work ersplant today, how many trees did? TÞÑ15trees originally, there are.21trees aftersome more were planted, then there were. Been 2115=6,sothere must have. VAÞÑ6 qPrompt 6 åQÞÑ32choco lates andhersisterhad42,Leah had. 35,ifthey ate,they have leftintotal,how many pieces do? TÞÑOriginally, 32choco lates, Leah had. 42,hersisterhad. 32+42=74,sointotalthey had. After eating35,74 -35=39,they had. VAÞÑ39 qPrompt 7 åQÞÑ$23, Olivia has.Five bagels for$3each, shebought. Shehave left,how much money does?

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

369

qPrompt 7 åQÞÑ$23, Olivia has.Five bagels for$3each, shebought. Shehave left,how much money does? TÞÑ23dollars, Olivia had. 5x3=15dollars, 5bagels for3dollarseach willbe.23-15dollarsleft, soshehas.23-15is8. VAÞÑ8 qPrompt 8 åQÞÑ58golf balls, Michael had. OnTues day,23 golf balls, helost. OnWednes day,2 more, helost. He have attheendofWednes day, how many golfballs did? TÞÑWith 58golfballs, Michael started. Afterlosing23onTues day,58 -23=35,hehad. Afterlosing 2more,35 -2=33golfballs, hehad. VAÞÑ33104 Table 74: S PORTS with examples in YodaSpeak. qPrompt 1 åQÞÑIsthefollowingsentence plausible? “Perfectfrom thelineJamalMurraywas.” TÞÑAbasketballplayer JamalMurrayis.Perfectfrom thelineispartofbasketballbeing.

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

370

TÞÑAbasketballplayer JamalMurrayis.Perfectfrom thelineispartofbasketballbeing. VAÞÑyes qPrompt 2 åQÞÑIsthefollowingsentence plausible? “The screen pass intheNFC cham pionship Joao Moutinho caught.” TÞÑAsoccerplayer Joao Moutinho is.PartofAmer icanfootballtheNFC cham pionship is,notsoccer. VAÞÑno qPrompt 3 åQÞÑIsthefollowingsentence plausible? “The buzzer Jonas Valan ciunasbeat.” TÞÑAbasketballplayer Jonas Valan ciunasis.Thebuzzer ispartofbasketballbeating. VAÞÑyes qPrompt 4 åQÞÑIsthefollowingsentence plausible? “Passed thepuck, Sam Darnold did.” TÞÑAnAmer icanfootballplayer Sam Darnold is.Part ofhockey passingthepuck is,notAmer ican football. VAÞÑno qPrompt 5

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

371

football. VAÞÑno qPrompt 5 åQÞÑIsthefollowingsentence plausible? “Called forslash ingKyle Palmieri was.” TÞÑAhockey player Kyle Palmieri is.Called forslash ingispartofhockey being. VAÞÑyes qPrompt 6 åQÞÑIsthefollowingsentence plausible? “The pick androllCarsonWentz set.” TÞÑAnAmer icanfootballplayer CarsonWentz is.Partofbasketballpick androllis,notfootball. VAÞÑno qPrompt 7 åQÞÑisthefollowingsentence plausible? In“Mal colm Brog donbanked theshot.” TÞÑAbasketballplayer Malcolm Brog donis.Inispartofbasketballbank ingtheshot. VAÞÑyes qPrompt 8 åQÞÑisthefollowingsentence plausible? Atouch down “Dray mond Green threw.” TÞÑAnbasketballplayer Dray mond Green is.Atouch down ispartoffootballthrow ing,notbasketball. VAÞÑno 105

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

372

TÞÑAnbasketballplayer Dray mond Green is.Atouch down ispartoffootballthrow ing,notbasketball. VAÞÑno 105 Table 75: D ATEwith examples in YodaSpeak. qPrompt 1 åQÞÑ4/19/1969 todayitis.Thedate 24hours later inMM/D D/YYYY what is? TÞÑ04/19/1969, todayis.Later isonedayaftertoday, 24hours, 04/20/1969, which would be. VAÞÑ04/20/1969 qPrompt 2 åQÞÑSched uled tobeon06/01/1943 theconcertwas, delayed byonedaytotodaybutwas. Thedate 10days ago inMM/D D/YYYY what is? TÞÑ06/02/1943, onedayafter06/01/1943 is,06/02/1943, sotodayis.05/23/1943, 10days beforetodayis. VAÞÑ05/23/1943 qPrompt 3

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

373

VAÞÑ05/23/1943 qPrompt 3 åQÞÑATues daythefirstdayof2019 is,thefirstMon dayof2019 andtodayis.Thedate todayinMM/D D/YYYY what is? TÞÑTues day, ifthefirstdayof2019 was, aTues day, then 01/01/2019 was. ThefirstMon day, todayis,sixdays later, would be.01/07/2019, sotodayis. VAÞÑ01/07/2019 qPrompt 4 åQÞÑBorn onthelastdayofFebru aryin2001 Jane was. Her16-year-oldbirth daytodayis.Thedate yesterdayin MM/D D/YYYY what is? TÞÑThe 28th, thelastdayofFebru aryis,Born on02/28/2001, soJane was. Her16-year oldbirth day, today is,02/28/2017, sotodayis.02/27/2017, soyesterdaywas. VAÞÑ02/27/2017 qPrompt 5

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

374

is,02/28/2017, sotodayis.02/27/2017, soyesterdaywas. VAÞÑ02/27/2017 qPrompt 5 åQÞÑCom ingin36hours 2015 is.Thedate oneweek from todayinMM/D D/YYYY what is?Yes. TÞÑCom ingin36hours, if2015 is,comingin2days, then itis.12/30/2014, 2days before 01/01/2015 is, 12/30/2014, sotodayis.01/05/2015, sooneweek from todaywillbe. VAÞÑ01/05/2015 qPrompt 6 åQÞÑTodayis3/11/2002 Jane thought, infactmar12buttodayis,1daylater which is.Thedate 24hours later in MM/D D/YYYY what is? TÞÑ03/12/2002, todayis.Later willbe03/13/2002, sothedate 24hours. VAÞÑ03/13/2002 106 Table 76: GSM-8 Kwith random thoughts. qPrompt 1

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

375

VAÞÑ03/13/2002 106 Table 76: GSM-8 Kwith random thoughts. qPrompt 1 åQÞÑShawn hasfivetoys. ForChrist mas, hegottwotoys each from hismom anddad. How many toys does he have now? TÞÑInitially, Steve is66inches tall.Aftergrow ing6inches, Steve is66+6=72inches tall. VAÞÑ9 qPrompt 2 åQÞÑIfthere are3carsintheparkinglotand2more carsarrive, how many carsareintheparkinglot? TÞÑSandraeats3beignets everymorn ingandthere are7days inaweek sosheeats3*7=21beignets ina week. VAÞÑ5 qPrompt 3 åQÞÑJasonhad20lollipops. Hegave Denny some lollipops. Now Jasonhas12lollipops. How many lollipops did Jasongive toDenny? TÞÑThequarterofthenumberis1,thus thenumberis1*4=4. VAÞÑ8 qPrompt 4

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

376

TÞÑThequarterofthenumberis1,thus thenumberis1*4=4. VAÞÑ8 qPrompt 4 åQÞÑThere were nine computersintheserver room. Five more computerswere installed each day, from Mon day toThurs day. How many computersarenow intheserver room? TÞÑHalf of10is10/2=5.Five more than 5is5+5=10. VAÞÑ29 qPrompt 5 åQÞÑThere are15trees inthegrove. Grove work erswillplant trees inthegrove today. Afterthey aredone, there willbe21trees. How many trees didthegrove work ersplant today? TÞÑAngie’s ageminus4is20-4=16.Thus, Angie’s ageis16. VAÞÑ6 qPrompt 6 åQÞÑLeah had32choco lates andhersisterhad42.Ifthey ate35,how many pieces dothey have leftintotal? TÞÑAtotalof20passengers areintwobuses. Thus, onebuscanfit20passengers /2buses =10passengers.

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

377

TÞÑAtotalof20passengers areintwobuses. Thus, onebuscanfit20passengers /2buses =10passengers. VAÞÑ39 qPrompt 7 åQÞÑOlivia has$23. Shebought fivebagels for$3each. How much money does shehave left? TÞÑ2bags weight 2pounds, soonebagweights 2pounds /2=1pound. So8bags oforanges would weigh 8* 1pound =8pounds. VAÞÑ8 qPrompt 8 åQÞÑMichael had58golfballs. OnTues day, helost23golfballs. OnWednes day, helost2more. How many golf balls didhehave attheendofWednes day? TÞÑJohn takes 2pills aday. Inaweek, John there foretakes 2*7=14pills. VAÞÑ33 107 Table 77: S PORTS with random thoughts. qPrompt 1 åQÞÑIsthefollowingsentence plausible? “JamalMurraywasperfectfrom theline.”

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

378

qPrompt 1 åQÞÑIsthefollowingsentence plausible? “JamalMurraywasperfectfrom theline.” TÞÑSam Darnold isaAmer icanfootballplayer. Passingthepuck ispartofhockey, notAmer icanfootball. VAÞÑyes qPrompt 2 åQÞÑIsthefollowingsentence plausible? “Joao Moutinho caught thescreen pass intheNFC cham pionship.” TÞÑDray mond Green isanbasketballplayer. Throw ingatouch down ispartoffootball, notbasketball. VAÞÑno qPrompt 3 åQÞÑIsthefollowingsentence plausible? “Jonas Valan ciunasbeat thebuzzer.” TÞÑCarsonWentz isanAmer icanfootballplayer. Pick androllispartofbasketball, notfootball. VAÞÑyes qPrompt 4 åQÞÑIsthefollowingsentence plausible? “Sam Darnold passed thepuck.” TÞÑJamalMurrayisabasketballplayer. Beingperfectfrom thelineispartofbasketball.

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

379

TÞÑJamalMurrayisabasketballplayer. Beingperfectfrom thelineispartofbasketball. VAÞÑno qPrompt 5 åQÞÑIsthefollowingsentence plausible? “Kyle Palmieri wascalled forslash ing.” TÞÑJoao Moutinho isasoccerplayer. TheNFC cham pionship ispartofAmer icanfootball, notsoccer. VAÞÑyes qPrompt 6 åQÞÑIsthefollowingsentence plausible? “Car sonWentz setthepick androll.” TÞÑMalcolm Brog donisabasketballplayer. Bank ingtheshot inispartofbasketball. VAÞÑno qPrompt 7 åQÞÑIsthefollowingsentence plausible? “Mal colm Brog donbanked theshot in.” TÞÑJonas Valan ciunasisabasketballplayer. Beatingthebuzzer ispartofbasketball. VAÞÑyes qPrompt 8 åQÞÑIsthefollowingsentence plausible? “Dray mond Green threw atouch down.”

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

380

qPrompt 8 åQÞÑIsthefollowingsentence plausible? “Dray mond Green threw atouch down.” TÞÑKyle Palmieri isahockey player. Beingcalled forslash ingispartofhockey. VAÞÑno 108 Table 78: D ATEwith random thoughts. qPrompt 1 åQÞÑItis4/19/1969 today. What isthedate 24hours later inMM/D D/YYYY? TÞÑThelastdayofFebru aryisthe28th, soJane wasborn on02/28/2001. Todayisher16-year oldbirth day, so todayis02/28/2017. Soyesterdaywas02/27/2017. VAÞÑ04/20/1969 qPrompt 2 åQÞÑTheconcertwassched uled tobeon06/01/1943, butwasdelayed byonedaytotoday. What isthedate 10 days agoinMM/D D/YYYY? TÞÑTodayis03/12/2002. Sothedate 24hours later willbe03/13/2002. VAÞÑ05/23/1943 qPrompt 3

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

381

VAÞÑ05/23/1943 qPrompt 3 åQÞÑThefirstdayof2019 isaTues day, andtodayisthefirstMon dayof2019. What isthedate todayinMM/D D/YYYY? TÞÑTodayis04/19/1969. 24hours later isonedayaftertoday, which would be04/20/1969. VAÞÑ01/07/2019 qPrompt 4 åQÞÑJane was born onthelastdayofFebru aryin2001. Todayisher16-year-oldbirth day. What isthedate yesterdayinMM/D D/YYYY? TÞÑOne dayafter06/01/1943 is06/02/1943, sotodayis06/02/1943. 10days beforetodayis05/23/1943. VAÞÑ02/27/2017 qPrompt 5 åQÞÑ2015 iscomingin36hours. What isthedate oneweek from todayinMM/D D/YYYY?

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

382

qPrompt 5 åQÞÑ2015 iscomingin36hours. What isthedate oneweek from todayinMM/D D/YYYY? TÞÑIfthefirstdayof2019 wasTues day, then 01/01/2019 wasaTues day. TodayisthefirstMon day, would be sixdays later. Sotodayis01/07/2019. VAÞÑ01/05/2015 qPrompt 6 åQÞÑJane thought todayis3/11/2002, buttodayisinfactMar 12,which is1daylater. What isthedate 24hours later inMM/D D/YYYY? TÞÑIf2015 iscomingin36hours, then itiscomingin2days. 2days before01/01/2015 is12/30/2014, sotoday is12/30/2014. Sooneweek from todaywillbe01/05/2015. VAÞÑ03/13/2002 109 Table 79: GSM-8 Kwith shuffled tokens within each sentence. The location of numbers is fixed. qPrompt 1

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

383

109 Table 79: GSM-8 Kwith shuffled tokens within each sentence. The location of numbers is fixed. qPrompt 1 åQÞÑShawn hasfivetoys. ForChrist mas, hegottwotoys each from hismom anddad. How many toys does he have now? TÞÑwith Shawn toys 5started. dad, from more 2histoys then istoys hemom gotthateach 4andIf.5+4=9 VAÞÑ9 qPrompt 2 åQÞÑIfthere are3carsintheparkinglotand2more carsarrive, how many carsareintheparkinglot? TÞÑoriginally carsThere 3are.2arrivemore cars. 3+2=5 VAÞÑ5 qPrompt 3 åQÞÑJasonhad20lollipops. Hegave Denny some lollipops. Now Jasonhas12lollipops. How many lollipops did Jasongive toDenny? TÞÑlollipops with started 20Jason. hadafterto12Denny Then some givinghe.hegave Denny So20-12=8 VAÞÑ8

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

384

VAÞÑ8 qPrompt 4 åQÞÑThere were nine computersintheserver room. Five more computerswere installed each day, from monday tothurs day. How many computersarenow intheserver room? TÞÑwere There originally 9computers.more Foreach 4computers5ofadded days, were. computers5*4=20 were added So.9+20is29 VAÞÑ29 qPrompt 5 åQÞÑThere are15trees inthegrove. Grove work erswillplant trees inthegrove today. Afterthey aredone, there willbe21trees. How many trees didthegrove work ersplant today? TÞÑareThere 15originally trees. planted were some 21more Then afterthere trees were. must Sothere been have 21-15=6 VAÞÑ6 qPrompt 6 åQÞÑLeah had32choco lates andhersisterhad42.Ifthey ate35,how many pieces dothey have leftintotal? TÞÑOriginally, hadchoco lates 32Leah. Hersisterhad42.totalhadthey inSo32+42=74.eatinghad35,After they 74-35=39

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

385

they 74-35=39 VAÞÑ39 qPrompt 7 åQÞÑOlivia has$23. Shebought fivebagels for$3each. How much money does shehave left? TÞÑhadOlivia 23dollars. 5dollarsbe3each dollarsbagels for5x3=15will. dollarsSoshe23-15hasleft.23 -15is8 VAÞÑ8 qPrompt 8 åQÞÑMichael had58golfballs. OnTues day, helost23golfballs. OnWednes day, helost2more. How many golf balls didhehave attheendofWednes day? TÞÑMichael started balls 58with golf. hadAfter23losingTues day,heon58-23=35.golflosing2balls more, he35-2=33Afterhad VAÞÑ33 110 Table 80: S PORTS with shuffled tokens within each sentence. qPrompt 1 åQÞÑIsthefollowingsentence plausible? “JamalMurraywasperfectfrom theline.” TÞÑisaplayer JamalbasketballMurray.from lineBeingperfectpartisthebasketballof

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

386

TÞÑisaplayer JamalbasketballMurray.from lineBeingperfectpartisthebasketballof VAÞÑyes qPrompt 2 åQÞÑIsthefollowingsentence plausible? “Joao Moutinho caught thescreen pass intheNFC cham pionship.” TÞÑMoutinho player issocceraJoao. NFC Amer icansoccerofcham pionship Thenotispartfootball, VAÞÑno qPrompt 3 åQÞÑIsthefollowingsentence plausible? “Jonas Valan ciunasbeat thebuzzer.” TÞÑisaValan ciunasJonas basketballplayer. partbuzzer basketballtheBeatingofis VAÞÑyes qPrompt 4 åQÞÑIsthefollowingsentence plausible? “Sam Darnold passed thepuck.” TÞÑaSam player footballisDarnold Amer ican. ofhockey, puck thefootballAmer icanisPassingpartnot VAÞÑno qPrompt 5 åQÞÑIsthefollowingsentence plausible? “Kyle Palmieri wascalled forslash ing.”

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

387

qPrompt 5 åQÞÑIsthefollowingsentence plausible? “Kyle Palmieri wascalled forslash ing.” TÞÑPalmieri player hockey aisKyle. called Beingofpartslash inghockey foris VAÞÑyes qPrompt 6 åQÞÑIsthefollowingsentence plausible? “Car sonWentz setthepick androll.” TÞÑfootballCarsonisAmer icanplayer Wentz an.rollandnotbasketball, partisPick footballof VAÞÑno qPrompt 7 åQÞÑIsthefollowingsentence plausible? “Mal colm Brog donbanked theshot in.” TÞÑplayer basketballBrog donMalcolm ais.theBank ingshot inofbasketballpartis VAÞÑyes qPrompt 8 åQÞÑIsthefollowingsentence plausible? “Dray mond Green threw atouch down.” TÞÑanisDray mond player basketballGreen. Throw ingfootball, aoftouch down partnotisbasketball VAÞÑno 111

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

388

TÞÑanisDray mond player basketballGreen. Throw ingfootball, aoftouch down partnotisbasketball VAÞÑno 111 Table 81: D ATEwith shuffled tokens of each sentence. qPrompt 1 åQÞÑItis4/19/1969 today. What isthedate 24hours later inMM/D D/YYYY? TÞÑToday04/19/1969 is.24afteronewhich later bewould hours 04/20/1969 daytoday, is VAÞÑ04/20/1969 qPrompt 2 åQÞÑTheconcertwassched uled tobeon06/01/1943, butwasdelayed byonedaytotoday. What isthedate 10 days agoinMM/D D/YYYY? TÞÑis06/02/1943, sotoday06/02/1943 day06/01/1943 One isafter.10todaydays isbefore05/23/1943 VAÞÑ05/23/1943 qPrompt 3

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

389

VAÞÑ05/23/1943 qPrompt 3 åQÞÑThefirstdayof2019 isaTues day, andtodayisthefirstMon dayof2019. What isthedate todayinMM/D D/YYYY? TÞÑwasTues daythedaywas2019 firstIfthen 01/01/2019 ofTues day, a.days Todaysixlater isMon day, be would firstthe.todaySo01/07/2019 is VAÞÑ01/07/2019 qPrompt 4 åQÞÑJane was born onthelastdayofFebru aryin2001. Todayisher16-year-oldbirth day. What isthedate yesterdayinMM/D D/YYYY? TÞÑlastdayJane Febru aryonwas The of28th, born issothe02/28/2001. Todaytodayisbirth day, oldso 02/28/2017 16-year isher.02/27/2017 yesterdaySowas VAÞÑ02/27/2017 qPrompt 5

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

390

02/28/2017 16-year isher.02/27/2017 yesterdaySowas VAÞÑ02/27/2017 qPrompt 5 åQÞÑ2015 iscomingin36hours. What isthedate oneweek from todayinMM/D D/YYYY? TÞÑdays 2015 isthen coming36Ifithours, comingisin2in.2is12/30/2014, days 01/01/2015 so12/30/2014 istodaybefore. from week willSo01/05/2015 betodayone VAÞÑ01/05/2015 qPrompt 6 åQÞÑJane thought todayis3/11/2002, buttodayisinfactMar 12,which is1daylater. What isthedate 24hours later inMM/D D/YYYY? TÞÑ03/12/2002 Todayis.thehours later 24bewillSodate 03/13/2002 VAÞÑ03/13/2002 112 Table 82: GSM-8 Kwith sentences within each thought are shuffled. The location of numbers is fixed. qPrompt 1

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

391

112 Table 82: GSM-8 Kwith sentences within each thought are shuffled. The location of numbers is fixed. qPrompt 1 åQÞÑShawn hasfivetoys. ForChrist mas, hegottwotoys each from hismom anddad. How many toys does he have now? TÞÑdad, hegot5toys. then started mom 2each ismore thatfrom Ifhisandtoys. toys 4with Shawn 5+4=9. VAÞÑ9 qPrompt 2 åQÞÑIfthere are3carsintheparkinglotand2more carsarrive, how many carsareintheparkinglot? TÞÑarrive. more are3cars2originally cars. There 3+2=5. VAÞÑ5 qPrompt 3 åQÞÑJasonhad20lollipops. Hegave Denny some lollipops. Now Jasonhas12lollipops. How many lollipops did Jasongive toDenny? TÞÑgivingSostarted 20some gave toDenny 12JasonDenny. Then afterhadhewith helollipops. 20-12=8.

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

392

TÞÑgivingSostarted 20some gave toDenny 12JasonDenny. Then afterhadhewith helollipops. 20-12=8. VAÞÑ8 qPrompt 4 åQÞÑThere were nine computersintheserver room. Five more computerswere installed each day, from Mon day toThurs day. How many computersarenow intheserver room? TÞÑoriginally were were 9computersadded. isFor4each 5computers.added. ofdays, more 5*4=20computerswere There 9+20So29. VAÞÑ29 qPrompt 5 åQÞÑThere are15trees inthegrove. Grove work erswillplant trees inthegrove today. Afterthey aredone, there willbe21trees. How many trees didthegrove work ersplant today? TÞÑSowere 15there aftermore have planted. 21trees originally. must There arethere were trees some Then been 21-15=6. VAÞÑ6 qPrompt 6 åQÞÑLeah had32choco lates andhersisterhad42.Ifthey ate35,how many pieces dothey have leftintotal?

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

393

åQÞÑLeah had32choco lates andhersisterhad42.Ifthey ate35,how many pieces dothey have leftintotal? TÞÑAfterOriginally, total32they sisterhadhad42.hadeatinginHerSo32+42=74.hadthey Leah 35, choco lates. 74-35=39. VAÞÑ39 qPrompt 7 åQÞÑOlivia has$23. Shebought fivebagels for$3each. How much money does shehave left? TÞÑbagels dollars23has5Olivia will3isdollars. left.for5had3=15dollars. Sobex23-15dollarseach 2315she8. VAÞÑ8 qPrompt 8 åQÞÑMichael had58golfballs. OnTues day, helost23golfballs. OnWednes day, helost2more. How many golf balls didhehave attheendofWednes day? TÞÑhegolfMichael 58golfAfterstarted with 23more, losinghadballs. 58-23=35.losingAfter2onTues day, he35-2=33balls. had VAÞÑ33 113

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

394

he35-2=33balls. had VAÞÑ33 113 Table 83: S PORTS with sentences within each thought are shuffled. qPrompt 1 åQÞÑIsthefollowingsentence plausible? “JamalMurraywasperfectfrom theline.” TÞÑperfectplayer. basketball. aMurraypartbasketballisofBeingJamalisthefrom line VAÞÑyes qPrompt 2 åQÞÑIsthefollowingsentence plausible? “Joao Moutinho caught thescreen pass intheNFC cham pionship.” TÞÑAmer icansoccer.Moutinho notsoccerJoao ispartplayer. NFC Thefootball, acham pionship is of VAÞÑno qPrompt 3 åQÞÑIsthefollowingsentence plausible? “Jonas Valan ciunasbeat thebuzzer.” TÞÑValan ciunasofplayer. thebasketballJonas Beatingisbuzzer ispartabasketball. VAÞÑyes qPrompt 4

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

395

VAÞÑyes qPrompt 4 åQÞÑIsthefollowingsentence plausible? “Sam Darnold passed thepuck.” TÞÑfootballispuck thehockey, player. notfootball. Darnold partaAmer icanofAmer icanPassing Sam is VAÞÑno qPrompt 5 åQÞÑIsthefollowingsentence plausible? “Kyle Palmieri wascalled forslash ing.” TÞÑhockey. ofPalmieri slash ingKyle Beingaplayer. called parthockey isisfor VAÞÑyes qPrompt 6 åQÞÑIsthefollowingsentence plausible? “Car sonWentz setthepick androll.” TÞÑisfootballnotfootball. CarsonPick rollbasketball, andpartWentz Amer icananofplayer. is VAÞÑno qPrompt 7 åQÞÑIsthefollowingsentence plausible? “Mal colm Brog donbanked theshot in.” TÞÑplayer. isofainBrog donbasketballBank ingshot thebasketball. partMalcolm is VAÞÑyes

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

396

TÞÑplayer. isofainBrog donbasketballBank ingshot thebasketball. partMalcolm is VAÞÑyes qPrompt 8 åQÞÑIsthefollowingsentence plausible? “Dray mond Green threw atouch down.” TÞÑofanaistouch down football, Green basketballDray mond basketball. notplayer. ispartThrow ing VAÞÑno 114 Table 84: D ATEwith sentences within each thought are shuffled. qPrompt 1 åQÞÑItis4/19/1969 today. What isthedate 24hours later inMM/D D/YYYY? TÞÑlater is04/19/1969. 24dayafterwhich would betoday, hours Today04/20/1969. oneis VAÞÑ04/20/1969 qPrompt 2 åQÞÑTheconcertwassched uled tobeon06/01/1943, butwasdelayed byonedaytotoday.What isthe date 10days agoinMM/D D/YYYY?

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

397

date 10days agoinMM/D D/YYYY? TÞÑtodaytodaydays One 05/23/1943. 06/01/1943 06/02/1943, 06/02/1943. isis10issodaybefore after VAÞÑ05/23/1943 qPrompt 3 åQÞÑThefirstdayof2019 isaTues day, andtodayisthefirstMon dayof2019. What isthedate today inMM/D D/YYYY? TÞÑ01/01/2019 todayMon day, would Tues day, 2019 first isTues day. dayoflater. then Sowas 01/07/2019. firstbedays TodaytheasixwasisIfthe VAÞÑ01/07/2019 qPrompt 4 åQÞÑJane wasborn onthelastdayofFebru aryin2001. Todayisher16-year-oldbirth day. What isthe date yesterdayinMM/D D/YYYY? TÞÑlastwasher16-year is02/28/2001. Soofso02/27/2017. isbirth day,isTheFebru aryonyesterday

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

398

TÞÑlastwasher16-year is02/28/2001. Soofso02/27/2017. isbirth day,isTheFebru aryonyesterday wasoldtodayToday02/28/2017. sodayborn 28th, theJane VAÞÑ02/27/2017 qPrompt 5 åQÞÑ2015 iscomingin36hours. What isthedate oneweek from todayinMM/D D/YYYY? TÞÑin2015 onehours, 01/01/2015 36soitdays. today12/30/2014, then 2week 2is12/30/2014. is bedays inis01/05/2015. Sofrom todaycomingcomingbeforeiswillIf VAÞÑ01/05/2015 qPrompt 6 åQÞÑJane thought todayis3/11/2002, buttodayisinfactMar 12,which is1daylater. What isthedate 24hours later inMM/D D/YYYY? TÞÑlater Sodate Today03/13/2002. hours 24is03/12/2002. thebewill VAÞÑ03/13/2002 115 Table 85: GSM-8 Kwith questions in YodaSpeak. qPrompt 1

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

399

VAÞÑ03/13/2002 115 Table 85: GSM-8 Kwith questions in YodaSpeak. qPrompt 1 åQÞÑFive toys, Shawn has.For Christ mas,Two toys each from hismom anddad, hegot.He have now, how many toys does? TÞÑShawn started with 5toys. Ifhegot2toys each from hismom anddad, then thatis4more toys. 5+4=9. VAÞÑ9 qPrompt 2 åQÞÑ3carsintheparkinglotand2more carsarrive, ifthere are,In theparkinglot,how many carsare? TÞÑThere areoriginally 3cars. 2more carsarrive. 3+2=5. VAÞÑ5 qPrompt 3 åQÞÑ20lollipops, Jasonhad. Denny some lollipops, hegave. 12lollipops, now Jasonhas. Jasongive toDenny, how many lollipops did? TÞÑJasonstarted with 20lollipops. Then hehad12aftergivingsome toDenny. Sohegave Denny 20-12=8.

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

400

TÞÑJasonstarted with 20lollipops. Then hehad12aftergivingsome toDenny. Sohegave Denny 20-12=8. VAÞÑ8 qPrompt 4 åQÞÑNine computersintheserver room, there were. Installed each day, fivemore computerswere, from Mon day toThurs day. Now intheserver room, how many computersare? TÞÑThere were originally 9computers.Foreach of4days, 5more computerswere added. So5*4=20 computerswere added. 9+20is29. VAÞÑ29 qPrompt 5 åQÞÑ15trees inthegrove, there are.Grove work erswill plant trees inthegrove today. Done, afterthey are,21 trees, there willbe.Thegrove work ersplant today, how many trees did? TÞÑThere are15trees originally. Then there were 21trees aftersome more were planted. Sothere must have been 21-15=6. VAÞÑ6 qPrompt 6

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

401

been 21-15=6. VAÞÑ6 qPrompt 6 åQÞÑ32choco lates andhersisterhad42,Leah had. 35,ifthey ate,they have leftintotal,how many pieces do? TÞÑOriginally, Leah had32choco lates. Hersisterhad42.Sointotalthey had32+42=74.Aftereating35,they had74-35=39. VAÞÑ39 qPrompt 7 åQÞÑ$23, Olivia has.Five bagels for$3each, shebought. Shehave left,how much money does? TÞÑOlivia had23dollars. 5bagels for3dollarseach willbe5x3=15dollars. Soshehas23-15dollarsleft.23 -15is8. VAÞÑ8 qPrompt 8 åQÞÑ58golfballs, Michael had. OnTues day,23golfballs, helost. OnWednes day,2more, helost. Hehave atthe endofWednes day, how many golfballs did?

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

402

endofWednes day, how many golfballs did? TÞÑMichael started with 58golfballs. Afterlosing23onTues day, hehad58-23=35.Afterlosing2more, he had35-2=33golfballs. VAÞÑ33 116 Table 86: S ORTING with verbose chain of thought. qPrompt 1 åQÞÑ7,8,4,1,2,9,3,6,5 TÞÑFirst numberis1.Next wehave 2,since 1islessthan 2.Next wehave 3,since 2islessthan 3.Next wehave 4,since 3islessthan 4.Next wehave 5,since 4islessthan 5.Next wehave 6,since 5islessthan 6.Next wehave 7, since 6islessthan 7.Next wehave 8,since 7islessthan 8.Next wehave 9,since 8islessthan 9. VAÞÑ1,2,3,4,5,6,7,8,9 qPrompt 2 åQÞÑ5,9,3,1,8,4,6,2

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

403

qPrompt 2 åQÞÑ5,9,3,1,8,4,6,2 TÞÑFirst numberis1.Next wehave 2,since 1islessthan 2.Next wehave 3,since 2islessthan 3.Next wehave 4,since 3islessthan 4.Next wehave 5,since 4islessthan 5.Next wehave 6,since 5islessthan 6.Next wehave 8, since 6islessthan 8.Next wehave 9,since 8islessthan 9. VAÞÑ1,2,3,4,5,6,8,9 qPrompt 3 åQÞÑ6,5,7,4,3,2,8,1 TÞÑFirst numberis1.Next wehave 2,since 1islessthan 2.Next wehave 3,since 2islessthan 3.Next wehave 4,since 3islessthan 4.Next wehave 5,since 4islessthan 5.Next wehave 6,since 5islessthan 6.Next wehave 7, since 6islessthan 7.Next wehave 8,since 7islessthan 8.

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

404

since 6islessthan 7.Next wehave 8,since 7islessthan 8. VAÞÑ1,2,3,4,5,6,7,8 qPrompt 4 åQÞÑ1,6,4,8,5,3,7,2 TÞÑFirst numberis1.Next wehave 2,since 1islessthan 2.Next wehave 3,since 2islessthan 3.Next wehave 4,since 3islessthan 4.Next wehave 5,since 4islessthan 5.Next wehave 6,since 5islessthan 6.Next wehave 7, since 6islessthan 7.Next wehave 8,since 7islessthan 8. VAÞÑ1,2,3,4,5,6,7,8 qPrompt 5 åQÞÑ5,2,1,4,3,7 TÞÑFirst numberis1.Next wehave 2,since 1islessthan 2.Next wehave 3,since 2islessthan 3.Next wehave 4,since 3islessthan 4.Next wehave 5,since 4islessthan 5.Next wehave 7,since 5islessthan 7.

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

405

4,since 3islessthan 4.Next wehave 5,since 4islessthan 5.Next wehave 7,since 5islessthan 7. VAÞÑ1,2,3,4,5,7 qPrompt 6 åQÞÑ3,8,2,5,6,4,7,1 TÞÑFirst numberis1.Next wehave 2,since 1islessthan 2.Next wehave 3,since 2islessthan 3.Next wehave 4,since 3islessthan 4.Next wehave 5,since 4islessthan 5.Next wehave 6,since 5islessthan 6.Next wehave 7, since 6islessthan 7.Next wehave 8,since 7islessthan 8. VAÞÑ1,2,3,4,5,6,7,8 qPrompt 7 åQÞÑ8,6,1,2,9,7,4 TÞÑFirst numberis1.Next wehave 2,since 1islessthan 2.Next wehave 4,since 2islessthan 4.Next wehave

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

406

TÞÑFirst numberis1.Next wehave 2,since 1islessthan 2.Next wehave 4,since 2islessthan 4.Next wehave 6,since 4islessthan 6.Next wehave 7,since 6islessthan 7.Next wehave 8,since 7islessthan 8.Next wehave 9, since 8islessthan 9. VAÞÑ1,2,4,6,7,8,9 qPrompt 8 åQÞÑ7,6,8,1 TÞÑFirst numberis1.Next wehave 6,since 1islessthan 6.Next wehave 7,since 6islessthan 7.Next wehave 8,since 7islessthan 8. VAÞÑ1,6,7,8 117 Table 87: GSM-8 Kwith verbalized numbers. qPrompt 1 åQÞÑShawn hasfivetoys. ForChrist mas, hegottwotoys each from hismom anddad. How many toys does he have now? TÞÑShawn started with fivetoys. Ifhegottwotoys each from hismom anddad, then thatisfour more toys. five

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

407

+four =9. VAÞÑ9 qPrompt 2 åQÞÑIfthere arethree carsintheparkinglotandtwomore carsarrive, how many carsareintheparkinglot? TÞÑThere areoriginally three cars. twomore carsarrive. three +two=five. VAÞÑ5 qPrompt 3 åQÞÑJasonhadtwenty lollipops. Hegave Denny some lollipops. Now Jasonhastwelve lollipops. How many lollipops didJasongive toDenny? TÞÑJasonstarted with twenty lollipops. Then hehadtwelve aftergivingsome toDenny. Sohegave Denny twenty -twelve =8. VAÞÑ8 qPrompt 4 åQÞÑThere were nine computersintheserver room. Five more computerswere installed each day, from Mon day toThurs day. How many computersarenow intheserver room? TÞÑThere were originally nine computers.Foreach offour days, fivemore computerswere added. Sofive*four =twenty computerswere added. nine +twenty is29.

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

408

=twenty computerswere added. nine +twenty is29. VAÞÑ29 qPrompt 5 åQÞÑThere arefifteen trees inthegrove. Grove work erswillplant trees inthegrove today. Afterthey aredone, there willbetwenty onetrees. How many trees didthegrove work ersplant today? TÞÑThere arefifteen trees originally. Then there were twenty onetrees aftersome more were planted. Sothere must have been twenty one-fifteen =6. VAÞÑ6 qPrompt 6 åQÞÑLeah hadthirty twochoco lates andhersisterhad42.Ifthey atethirty five,how many pieces dothey have left intotal? TÞÑOriginally, Leah hadthirty twochoco lates. Hersisterhad42.Sointotalthey hadthirty two+forty two= 74.Aftereatingthirty five, they hadseventy four -thirty five=39. VAÞÑ39 qPrompt 7

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

409

VAÞÑ39 qPrompt 7 åQÞÑOlivia hastwenty three. Shebought fivebagels forthree dollarseach. How much money does shehave left? TÞÑOlivia hadtwenty three dollars. fivebagels forthree dollarseach willbefivexthree =fifteen dollars. Soshe hastwenty three -fifteen dollarsleft.twenty three -fifteen is8. VAÞÑ8 qPrompt 8 åQÞÑMichael hadfifty eight golf balls. OnTues day, helosttwenty three golf balls. OnWednes day, helosttwo more. How many golfballs didhehave attheendofWednes day? TÞÑMichael started with fifty eight golf balls. Afterlosingtwenty three onTues day, hehadfifty eight -twenty three =3five. Afterlosingtwomore, hehadthirty five-two=thirty three golfballs. VAÞÑ33 118 Table 88: S PORTS with abstract sports person, sport, and sport activity. qPrompt 1

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

410

VAÞÑ33 118 Table 88: S PORTS with abstract sports person, sport, and sport activity. qPrompt 1 åQÞÑIsthefollowingsentence plausible? "PER SON wasinvolved inACTIVITY ." TÞÑPERSON isaSPORT1 player. BeingACTIVITY ispartofSPORT1. VAÞÑyes qPrompt 2 åQÞÑIsthefollowingsentence plausible? "PER SON does ACTIVITY ." TÞÑPERSON isaSPORT2 player. TheACTIVITY ispartofSPORT3, notSPORT2. VAÞÑno qPrompt 3 åQÞÑIsthefollowingsentence plausible? "PER SON finishes ACTIVITY ." TÞÑPERSON isaSPORT1 player. ACTIVITY ispartofSPORT1. VAÞÑyes qPrompt 4 åQÞÑIsthefollowingsentence plausible? "PER SON performs ACTIVITY ." TÞÑPERSON isaSPORT3 player. ACTIVITY ispartofSPORT4, notSPORT3. VAÞÑno

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

411

TÞÑPERSON isaSPORT3 player. ACTIVITY ispartofSPORT4, notSPORT3. VAÞÑno qPrompt 5 åQÞÑIsthefollowingsentence plausible? "PER SON wasseen doingACTIVITY ." TÞÑPERSON isaSPORT4 player. BeingACTIVITY ispartofSPORT4. VAÞÑyes qPrompt 6 åQÞÑIsthefollowingsentence plausible? "PER SON complete ACTIVITY ." TÞÑPERSON isanSPORT3 player. ACTIVITY ispartofSPORT1, notSPORT3. VAÞÑno qPrompt 7 åQÞÑIsthefollowingsentence plausible? "PER SON executes ACTIVITY ." TÞÑPERSON isaSPORT1 player. ACTIVITY ispartofSPORT1. VAÞÑyes qPrompt 8 åQÞÑIsthefollowingsentence plausible? "PER SON carriesoutACTIVITY ." TÞÑPERSON isanSPORT1 player. ACTIVITY ispartofSPORT3, notSPORT1.

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

412

TÞÑPERSON isanSPORT1 player. ACTIVITY ispartofSPORT3, notSPORT1. VAÞÑno 119 Table 89: S PORTS with abstract sports person and sport. qPrompt 1 åQÞÑIsthefollowingsentence plausible? “PER SON wasperfectfrom theline.” TÞÑPERSON isaSPORT1 player. Beingperfectfrom thelineispartofSPORT1. VAÞÑyes qPrompt 2 åQÞÑIsthefollowingsentence plausible? “PER SON caught thescreen pass intheNFC cham pionship.” TÞÑPERSON isaSPORT3 player. TheNFC cham pionship ispartofSPORT2, notSPORT3. VAÞÑno qPrompt 3 åQÞÑIsthefollowingsentence plausible? “PER SON beat thebuzzer.” TÞÑPERSON isaSPORT1 player. Beatingthebuzzer ispartofSPORT1. VAÞÑyes qPrompt 4

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

413

TÞÑPERSON isaSPORT1 player. Beatingthebuzzer ispartofSPORT1. VAÞÑyes qPrompt 4 åQÞÑIsthefollowingsentence plausible? “PER SON passed thepuck.” TÞÑPERSON isaSPORT2 player. Passingthepuck ispartofSPORT4, notSPORT2. VAÞÑno qPrompt 5 åQÞÑIsthefollowingsentence plausible? “PER SON wascalled forslash ing.” TÞÑPERSON isaSPORT4 player. Beingcalled forslash ingispartofSPORT4. VAÞÑyes qPrompt 6 åQÞÑIsthefollowingsentence plausible? “PER SON setthepick androll.” TÞÑPERSON isanSPORT2 player. Pick androllispartofSPORT1, notSPORT2. VAÞÑno qPrompt 7 åQÞÑIsthefollowingsentence plausible? “PER SON banked theshot in.”

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

414

VAÞÑno qPrompt 7 åQÞÑIsthefollowingsentence plausible? “PER SON banked theshot in.” TÞÑPERSON isaSPORT1 player. Bank ingtheshot inispartofSPORT1. VAÞÑyes qPrompt 8 åQÞÑIsthefollowingsentence plausible? “PER SON threw atouch down.” TÞÑPERSON isanSPORT1 player. Throw ingatouch down ispartofSPORT2, notSPORT1. VAÞÑno 120 Table 90: S PORTS with the order of clauses switched. qPrompt 1 åQÞÑIsthefollowingsentence plausible? "Jamel Murraywasperfectfrom theline." TÞÑBeingperfectfrom thelineispartofbasketball. JamalMurrayisabasketballplayer. VAÞÑyes qPrompt 2 åQÞÑIsthefollowingsentence plausible? "Joao Moutinho caught thescreen pass intheNFC cham pionship."

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

415

qPrompt 2 åQÞÑIsthefollowingsentence plausible? "Joao Moutinho caught thescreen pass intheNFC cham pionship." TÞÑTheNFC cham pionship ispartofAmer icanfootball, notsoccer.Joao Moutinho isasoccerplayer. VAÞÑno qPrompt 3 åQÞÑIsthefollowingsentence plausible? "Jonas Valan ciunasbeat thebuzzer." TÞÑBeatingthebuzzer ispartofbasketball. Jonas Valan ciunasisabasketballplayer. VAÞÑyes qPrompt 4 åQÞÑIsthefollowingsentence plausible? "Sam Darnold passed thepuck." TÞÑPassingthepuck ispartofhockey, notAmer icanfootball. Sam Darnold isaAmer icanfootball player. VAÞÑno qPrompt 5 åQÞÑIsthefollowingsentence plausible? "Kyle Palmieri wascalled forslash ing." TÞÑBeingcalled forslash ingispartofhockey. Kyle Palmieri isahockey player.

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

416

TÞÑBeingcalled forslash ingispartofhockey. Kyle Palmieri isahockey player. VAÞÑyes qPrompt 6 åQÞÑIsthefollowingsentence plausible? "CarsonWentz setthepick androll." TÞÑPick androllispartofbasketball, notfootball. CarsonWentz isanAmer icanfootballplayer. VAÞÑno qPrompt 7 åQÞÑIsthefollowingsentence plausible? "Mal colm Brog donbanked theshot in." TÞÑBank ingtheshot inispartofbasketball. Malcolm Brog donisabasketballplayer. VAÞÑyes qPrompt 8 åQÞÑIsthefollowingsentence plausible? "Dray mond Green threw atouch down." TÞÑThrow ingatouch down ispartoffootball, notbasketball. Dray mond Green isanbasketballplayer. VAÞÑno 121

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

1811.11206

0

Partitioned Variational Inference: A unified framework encompassing federated and continual learning Thang D. Bui1, Cuong V. Nguyen2, Siddharth Swaroop2, and Richard E. Turner2 1University of Sydney, Australia; [email protected] 2University of Cambridge, UK; {vcn22,ss2163,ret26}@cam.ac.uk Abstract Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

1

on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI. 1 Introduction Variational methods recast approximate inference as an optimization problem, thereby enabling advances in optimization to be leveraged for inference. VI has enabled approaches including natural gradient methods, mirror-descent, trust region and stochastic (mini-batch) optimization to be tapped in this way. The approach has been successful, with VI methods often lying on the efficient frontier of approximate inference’s speed-accuracy trade-off. VI has consequently become one of the most popular varieties of approximate inference. For example, it is now a standard approach for Gaussian process models [Titsias,

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

2

approximate inference. For example, it is now a standard approach for Gaussian process models [Titsias, 2009], latent topic models [Blei et al., 2003], and deep generative models [Kingma and Welling, 2014]. Deployment of VI requires the practitioner to make three fundamental choices. First, the form of the approximate family which ranges from simple mean-field or factorized distributions, through unfactorized exponential families to complex non-exponential family distributions. Second, the granularity of variational inference which includes, on the one hand, approaches based on the global variational free-energy, and on the other those that consider a single data point at a time and employ local message passing. Third, the form of the variational updates which encompasses the optimization method employed for maximizing the global variational free-energy or the form of the message passing updates. A large body of work has investigated how the choice of approximating family affects the accuracy of VI [MacKay, 2003, Wang and Titterington, 2004, Turner and Sahani, 2011] and how additional approximations can enable VI to support more complex approximate families [Jaakkola and Jordan, 1arXiv:1811.11206v1 [stat.ML] 27 Nov 2018

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

3

1arXiv:1811.11206v1 [stat.ML] 27 Nov 2018 1998, Rezende and Mohamed, 2015, Salimans et al., 2015, Ranganath et al., 2016, Mescheder et al., 2017]. This is a fundamental question, but it is orthogonal to the focus of the current paper. Instead, we focus on the second two choices. The granularity of variational inference is an important algorithmic dimension. Whilst global variational inference has more theoretical guarantees and is arguably simpler to implement, local variational inference offers unique opportunities for online or continual learning (e.g. allowing ‘old’ data to be sporadically revisited) and distributed computing (e.g. supporting asynchronous lock-free updates). The form of the updates is equally important with a burgeoning set of alternatives. For global VI these including gradient ascent, natural gradient and mirror descent, approximate second-order methods, stochastic versions thereof, collapsed VI and fixed-point updates to name but a few. For local VI, there has been less exploration of the options, but damping in natural and moment space is often employed.

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

4

VI, there has been less exploration of the options, but damping in natural and moment space is often employed. The goal of this paper is to develop a unifying framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges that the granularity and the optimization method are two fundamental algorithmic dimensions of VI. The new framework 1. generalizes and extends current theoretical results in this area, 2. reveals the relationship between a large number of existing schemes, and 3. identifies opportunities for innovation, a selection of which are demonstrated in experiments. We briefly summarize the contributions of this paper, focusing on the unified viewpoint and novel algorithmic extensions to support federated and continual learning. 1.1 Unification The main theoretical contributions of the paper, described in sections 2 to 4, are: to develop Partitioned Variational Inference; clean up, generalize and derive new supporting theory (including PVI fixed-point optimization, mini-batch approximation, hyperparameter learning); and show that PVI subsumes standard global variational inference, (local) variational message passing, and other well-established approaches. In addition, we also show in section 4 that damped fixed-point optimization and natural

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

5

approaches. In addition, we also show in section 4 that damped fixed-point optimization and natural gradient methods applied to PVI are equivalent to variationally-limited power EP. In section 4 PVI is used to connect a large literature that has become fragmented with separated strandsofrelated, butmutuallyuncitedwork. Morespecificallyweunifyworkon: onlineVI[Ghahramani and Attias, 2000, Sato, 2001, Broderick et al., 2013, Bui et al., 2017b, Nguyen et al., 2018]; global VI [Sato, 2001, Hensman et al., 2012, Hoffman et al., 2013, Salimans and Knowles, 2013, Sheth and Khardon, 2016a, Sheth et al., 2015, Sheth and Khardon, 2016b]; local VI [Knowles and Minka, 2011, Wand, 2014, Khan and Lin, 2018]; power EP and related algorithms [Minka, 2001, 2004, Li et al., 2015, Hasenclever et al., 2017, Gelman et al., 2014]; and stochastic mini-batch variants of these algorithms [Hoffman et al., 2013, Li et al., 2015, Khan and Lin, 2018]. Figures 2 and 3 and table 1 present a

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

6

[Hoffman et al., 2013, Li et al., 2015, Khan and Lin, 2018]. Figures 2 and 3 and table 1 present a summary of these relationships in the context of PVI. 1.2 Probabilistic inference for federated machine learning The goal of federated learning is to enable distributed training of machine learning models without centralizing data [see e.g. McMahan et al., 2017, Zhao et al., 2018]. This is challenging in practice as: modern data sets can often be distributed inhomogeneously and unevenly across many machines, forexamples, mobiledevicescancontainmanyimageswhichcanbeusedfortrainingaclassification model, but accessing such information is often restricted and privacy-sensitive; computation resources available at terminal machines can be leveraged, but communication between these machines or between them and a central server can be limited and unreliable, for example, communication from and to mobile devices is often costly, and each device can be abruptly disconnected from the training setup or, similarly, a new device can appear; 2 the inference or prediction step is often needed in an any-time fashion at each machine, i.e. each machine needs to have access to a high-quality model to make predictions without having to send data to a remote server.

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

7

machine needs to have access to a high-quality model to make predictions without having to send data to a remote server. These requirements are often not satisfied in the traditional training pipelines, many of which require data to be stored in a single machine, or in a data center where it is typically distributed among many machines in a homogeneous and balanced fashion [see e.g. Dean et al., 2012, Zhang et al., 2015, Chen et al., 2016]. Federated learning attempts to bridge this gap by tackling the aforementioned constraints. Additionally, this type of learning is arguably less privacy-sensitive as compared to centralized learning approaches, as it does not require local data to be collected and sent to a central server. It can also be further improved by employing encrypted aggregation steps [Bonawitz et al., 2017] or differentially-private mechanisms [Dwork and Roth, 2014]. Distributed inference is also an active research area in the Bayesian statistics and machine learning literature. For example, parallel Markov chain Monte Carlo approaches typically run multiple independent Markov chains on different partitions of the data set, but require heuristics to aggregate, reweight and average the samples at test time [see e.g. Wang and Dunson, 2013, Scott et al., 2016].

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

8

reweight and average the samples at test time [see e.g. Wang and Dunson, 2013, Scott et al., 2016]. The closest to our work is the distributed EP algorithms of Gelman et al. [2014] and Hasenclever et al. [2017], which employ (approximate) MCMC for data partitions and EP for communication between workers. However, it is not clear these distributed approaches will work well in the federated settings described above. In section 5, we demonstrate that PVI can naturally and flexibly address the above challenges, and thus be used for federated learning with efficient synchronous or lock-free asynchronous communication. The proposed approach can be combined with recent advances in Monte Carlo VI for neural networks, enabling fast and communication-efficient training of Bayesian neural networks on non-iid federated data. We provide an extensive experiment comparing to alternative approaches in section 7. 1.3 Probabilistic inference for continual learning Continual learning (also termed online learning or life-long learning or incremental learning) is the ability to learn continually and adapt quickly to new experiences without catastrophically forgetting previously seen experiences [Schlimmer and Fisher, 1986, McCloskey and Cohen, 1989, Sutton and

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

9

ability to learn continually and adapt quickly to new experiences without catastrophically forgetting previously seen experiences [Schlimmer and Fisher, 1986, McCloskey and Cohen, 1989, Sutton and Whitehead, 1993, Ratcliff, 1990]. Such requirements arise in many practical settings in which data can arrive sequentially or tasks may change over time (e.g. new classes may be discovered), or entirely new tasks can emerge. Batch learning algorithms which deal with the entire data set at once are not applicable in these settings, as (1) data can arrive one point at a time or in batches of a size that is unknown a priori, or in a possibly non i.i.d. way; and (2) previously seen data may not be directly accessible, which means the continual learning algorithms need to intelligently decide how to best combine prior or current experience with new data while being resistant to under-fitting or over-fitting to new data (i.e. intransigence vs forgetting). Continual learning has a rich literature [see e.g. Opper, 1998, Sato, 2001, Ghahramani and Attias, 2000, Csató and Opper, 2002, Minka, 2001, Smola et al., 2004] but is enjoying a resurgence of interest

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

10

2000, Csató and Opper, 2002, Minka, 2001, Smola et al., 2004] but is enjoying a resurgence of interest ranging from deepening understanding of transfer learning and catastrophic forgetting [Goodfellow et al., 2014, Flesch et al., 2018], to developing learning algorithms for various models and applications [Broderick et al., 2013, Li and Hoiem, 2016, Kirkpatrick et al., 2017, Zenke et al., 2017, Seff et al., 2017, Bui et al., 2017a, Nguyen et al., 2018, Zeno et al., 2018, Chaudhry et al., 2018], to setting up relevant metrics and benchmarks for evaluation [Lomonaco and Maltoni, 2017, Hayes et al., 2018]. While the PVI framework enables us to connect and unify much of the literature in this area, it also allows gaps in the literature to be identified and enables the development of new and improved algorithmic solutions. We demonstrate this in section 6 by presenting a new continual learning method for Gaussian process regression and classification that greatly extends earlier work by Csató and Opper [2002] and Bui et al. 3 [2017a], allowing principled handling of hyperparameters and private pseudo-points for new data. The

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

11

3 [2017a], allowing principled handling of hyperparameters and private pseudo-points for new data. The new technique is shown to be superior to alternative online learning approaches on various toy and real-world data sets in section 7. We also show in section 5 that continual learning can be reframed as a special case of federated learning. 2 Partitioned Variational Inference In this section, we introduce Partitioned Variational Inference, a framework that encompasses many approaches to variational inference. We begin by framing PVI in terms of a series of local variational free-energy optimization problems, proving several key properties of the algorithm that reveal the relationship to global VI. In order to keep the development clear, we have separated most of the discussion of related work into section 4. Consider a parametric probabilistic model defined by the prior p(j)over parameters and the likelihood function p(yj;) =QM m=1p(ymj;), wherefy1;:::; yMgis a partition of yintoMgroups of data points. Depending on the context, a data group ymcan be considered to be a mini-batch of

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

12

of data points. Depending on the context, a data group ymcan be considered to be a mini-batch of ywhich is fixed across epochs, or a data shard. For simplicity, we assume for the moment that the hyperparameters are fixed and suppress them to lighten the notation. We will discuss hyperparameter optimization at the end of this section. Exact Bayesian inference in this class of model is in general intractable so we resort to variational inference. In particular, we posit a variational approximation of the true posterior as follows, q() =p()MY m=1tm()1 Zp()MY m=1p(ymj) =p(jy); (1) whereZis the normalizing constant of the true posterior, or marginal likelihood. The approximate likelihoodtm()will be refined by PVI to approximate the effect the likelihood term p(ymj)has on the posterior. Note that the form of q()in(1)is similar to that employed by the expectation propagation algorithm [Minka, 2001], but with two differences. First, the approximate posterior is not restricted to

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

13

algorithm [Minka, 2001], but with two differences. First, the approximate posterior is not restricted to lie in the exponential family, as is typically the case for EP. Second, the approximate posterior does not include a normalizing constant. Instead, the PVI algorithm will automatically ensure that the product of the prior and approximate likelihood factors in (1)is a normalized distribution. We will show that PVI will return an approximation to the marginal likelihood logZ=logp(y)in addition to the approximation of the posterior. Algorithm 1 details the PVI algorithm. At each iteration i, we select an approximate likelihood to refine according to a schedule bi2f1:::Mg. The approximate likelihood t(i1) bi()obtained from the previous iteration will be refined and the corresponding data-group is denoted ybi. The refinement proceeds in two steps. First, we refine the approximate posterior using the local (negative) variational free energy q(i)() = argmaxq()2QF(i)(q())where the optimization is over a tractable family Qand F(i)(q()) =Z

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

14

F(i)(q()) =Z dq() logq(i1)()p(ybij) q()t(i1) bi(): (2) Second, the new approximate likelihood is found by division, t(i) bi() =q(i)() q(i1)()t(i1) bi(). We will now justify these steps by stating properties, derived in the appendix, that show 1) the local free-energy optimization is equivalent to a variational KL optimization, 2) the update for the approximate likelihoods is consistent with the normalized density specified in 1, and 3) any fixed point of the algorithm is also a local optimum of global VI and at this fixed point the sum of the local 4 free-energies is equal to the global variational free-energy. The following properties apply for general q(), and are not limited to the exponential family.1 Property 1 Maximizing the local free-energy F(i)(q())is equivalent to the KL optimization q(i)() = argmin q()2QKL q()kbp(i)()

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

15

q(i)() = argmin q()2QKL q()kbp(i)() ; (3) wherebp(i)() =1 bZiq(i1)() t(i1) bi()p(ybij) =1 bZip(ybij)Q m6=bit(i1) m()is known as the tilted distribution in the EP literature and is intractable. The proof is straightforward (see A.1). The tilted distribution can be justified as a sensible target as it removes the approximate likelihood t(i1) bi()from the current approximate posterior and replaces it with the true likelihood p(ybij). In this way, the tilted distribution comprises one true likelihood, M1approximate likelihoods and the prior. The KL optimization then ensures the new posterior better approximates the true likelihood’s effect, in the context of the approximate likelihoods and the prior. Property 2 At the end of each iteration i= 0;1;:::,q(i)() =p()QM m=1t(i) m().

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

16

m=1t(i) m(). Again the proof is simple (see A.2), but it relies on PVI initializing the approximate likelihood factors to unity so that q(0)() =p(). Property 3 Letq() =p()QM m=1t m()be a fixed point of Algorithm 1, Fm(q()) =Rdq()logq()p(ymj) q()tm()be the local free-energy w.r.t. the factor tm, andF(q()) = R dq() logp()QM m=1p(ymj) q()be the global free-energy. We have: (a)PM m=1Fm(q()) =F(q()), i.e. the sum of the local free-energies is equal to the global free-energy, i.e. the PVI fixed point is an optimum of global VI,

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

17

free-energy, i.e. the PVI fixed point is an optimum of global VI, (b) Ifq() = argmaxq()2QFm(q())for allm, thenq() = argmaxq()2QF(q()). These results are more complex to show, but can be derived by computing the derivative and Hessian of the global free-energy and substituting into these expressions the derivatives and Hessians of the local free-energies (see A.3). The fact that the fixed point of PVI recovers a global VI solution (both the optimal q()and the global free-energy at this optimum) is the main theoretical justification for employing PVI. However, we do not believe that there is a Lyapunov function for PVI, indicating that it may oscillate or diverge in general. Having laid out the general framework for PVI, what remains to be decided is the method used for optimizing the local free-energies. In a moment we consider three choices: analytic updates, off-the-shelf optimization methods and fixed-point iterations, as well as discussing how stochastic approximations

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

18

optimization methods and fixed-point iterations, as well as discussing how stochastic approximations can be combined with these approaches. Before turning to these choices, we compare and contrast the algorithmic benefits of the local and global approaches to VI in different settings. This discussion will help shape the development of the optimization choices which follows. 2.1 When should a local VI approach be employed rather than a global one? We will describe in section 4 how the PVI framework unifies a large body of existing literature, thereby providing a useful conceptual scaffold for understanding the relationship between algorithms. However, 1However, we will only consider exponential family approximations in the experiments in section 7. 5 Algorithm 1 Partitioned Variational Inference Input:data partitionfy1;:::; yMg, priorp() Initialize: t(0) m():= 1for allm= 1;2;:::;M: q(0)():=p(): fori= 1;2;:::until convergence do bi:=index of the next approximate likelihood to refine. Compute the new approximate posterior: q(i)():= argmax q()2QZ

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

19

Compute the new approximate posterior: q(i)():= argmax q()2QZ dq() logq(i1)()p(ybij) q()t(i1) bi() Update the approximate likelihood: t(i) bi():=q(i)() q(i1)()t(i1) bi(); (4) t(i) m():=t(i1) m()for allm6=bi: end for Figure 1: Steps of the PVI algorithm when being used for continual learning [left] and federated learning [right]. it is important to ask: What algorithmic and computation benefits, if any, arise from considering a set of local free-energy updates, rather than a single global approximation (possibly leveraging stochastic mini-batch approximation)? In a nutshell, we will show that if the data set is fixed before inference is performed (batch learning) or arrives in a simple online iid way (simple online learning), and distributed computation is not available, 6 then global VI will typically be simpler to implement, require less memory, and faster to converge than

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

20

6 then global VI will typically be simpler to implement, require less memory, and faster to converge than more local versions of PVI (the case of scaling collapsed bounds being a possible exception). However, if the conditions above are not met, the local versions of PVI will be appropriate. We will now unpack important examples of this sort. The PVI approach is ideally suited to the distributed setting, with simple distributed variants allowing asynchronous distributed updates. One simple approach, similar to that of Hasenclever et al. [2017], uses Mworkers that are each allocated a data group ym. The workers store and refine the associated approximate likelihood tm(). A server maintains and updates the approximate posterior and communicates it to the workers. An idle worker receives the current posterior from the server, optimizes thelocalfree-energy, computesthechangeinthelocalapproximatelikelihood
m() =t(new) m()=t(old) m(), sends this to the server, and repeats. The local workers do not change q()directly. Instead, the server maintains a queue of approximate likelihood updates and applies these to the approximate posterior q(new)() =q(old)()
m(). This setup supports asynchronous updates of the approximate likelihood

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

21

q(new)() =q(old)()
m(). This setup supports asynchronous updates of the approximate likelihood factors. See fig. 1 for a pictorial depiction of these steps. In contrast, global VI is generally ill-suited to the distributed setting. Although the free-energy optimization can be parallelized over data points, typically this will only be advantageous for large mini-batches where the extra communication overhead does not dominate. Large mini-batches often result in slow optimization progress (early in learning it is often clear how to improve q()after seeing only a small number of data points). The special case of global VI employing mini-batch approximations and natural gradient updates can support asynchronous distributed processing if each worker receives statistically identical data and updates with the same frequency. It could not operate successfully when each node contains different amounts or types of data, or if some workers update more frequently than others. Distributed versions of PVI not only enable VI to be scaled to large problems, but they also allow inference algorithms to be sent to user data, rather than requiring user data to be collected and centralized before performing inference. Consider the situation where workers are personal devices, like mobile phones, containing user data ym. Here the local free-energy updates can be performed

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

22

centralized before performing inference. Consider the situation where workers are personal devices, like mobile phones, containing user data ym. Here the local free-energy updates can be performed client-side on the user’s devices and only summaries tm()of the relevant aspects of that information are communicated back to the central server. The frequency with which these messages are sent might be limited to improve security. Such an implementation is arguably more secure than one in which the user data (or associated gradients) are sent back to a central server [The Royal Society, 2017]. Since the amount and type of data at the nodes is outside of the control of the algorithm designer, mini-batch natural gradient global VI will generally be inappropriate for this setting. The PVI approach is also well suited to the continual or life-long learning setting. These settings are very general forms of online learning in which new data regularly arrive in a potentially non-iid way, tasks may change over time, and entirely new tasks may emerge. In this situation, the PVI framework can not only be used to continuously update the posterior distribution q()in light of new data by optimizing the local free-energy for the newly seen data, it can also be used to revisit old data groups (potentially in a judiciously selected way) thereby mitigating problems like catastrophic

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

23

data groups (potentially in a judiciously selected way) thereby mitigating problems like catastrophic forgetting. The update steps for this learning scenario are illustrated in fig. 1. In contrast, global VI is fundamentally ill-suited to the general online setting. The special case of global VI employing mini-batch approximations with natural gradient updates may be appropriate when the data are iid and only one update is performed for each new task (simple online learning), but it is not generally applicable. We will return to discuss the key issues raised in this section – the speed of convergence, memory overhead, online learning, and distributed inference – in the context of different options for carrying out the optimization of the local free-energies in section 3. 7 2.2 Hyperparameter Learning Many probabilistic models depend on a set of hyperparameters and it is often necessary to learn suitable settings from data to achieve good performance on a task. One method is to optimize the variational free-energy thereby approximating maximum likelihood learning. The gradient of the global variational free-energy decomposes into a set of local computations, as shown in appendix B, d dF(;q()) =MX m=1Eq()d

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

24

d dF(;q()) =MX m=1Eq()d dlogp(ymj;) +Eq()d dlogp(j) : (5) This expression holds for general q()and is valid both for coordinate ascent (updating withq() fixed) and for optimizing the collapsed bound (where the approximate posterior optimizes the global free-energy q() =q()and therefore depends implicitly on ). Notice that this expression is amenable to stochastic approximation which leads to optimization schemes that use only local information at each step. When combined with different choices for the optimization of the local free-energies wrt q(), this leads to a wealth of possible hyperparameter optimization schemes. In cases where a distributional estimate for the hyperparameters is necessary, e.g. in continual learning, the PVI framework above can be extended to handle the hyperparameters. In particular, the approximate posterior in eq. (1) can be modified as follows, q(;) =p()p(j)MY m=1tm(;)1

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

25

q(;) =p()p(j)MY m=1tm(;)1 Zp()p(j)MY m=1p(ymj;) =p(;jy);; (6) where the approximate likelihood factor tm(;)now involves both the model parameters and the hyperparameters. Similar to eq. (2), the approximate posterior above leads to the following local variational free-energy, F(i)(q(;)) =Z ddq(;) logq(i1)(;)p(ybij;) q(;)t(i1) bi(;): (7) Note that this approach retains all favourable properties of PVI such as local computation and flexibility in choosing optimization strategies and stochastic approximations. 3 Approaches for Optimizing the Local Free-energies Having established the general PVI algorithm and its properties, we will now describe different options for performing the optimization of the local free-energies. 3.1 Analytic Local Free-energy Updates

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

26

for performing the optimization of the local free-energies. 3.1 Analytic Local Free-energy Updates Each local free-energy is equivalent in form to a global free-energy with an effective prior pe() = q(i1)()=t(i1) bi(). As such, in conjugate exponential family models the KL optimizations will be available in closed form, for example in GP regression, and these updates can be substituted back into the local variational free-energies to yield locally-collapsed bounds, Fn(q(i)()), that are useful for hyperparameter optimization [Bui et al., 2017a]. One advantage of using local versions of PVI is that this allows collapsed bounds to be leveraged on large data sets where an application to entire data set would be computationally intractable, potentially speeding up convergence over global VI. 8 3.2 Off-the-shelf Optimizers for Local Free-energy Optimization If analytic updates are not tractable, the local free-energy optimizations can be carried out using standard optimizers. The PVI framework automatically breaks the data set into a series of local free-energy optimization problems and the propagation of uncertainty between the data groups weights

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

27

standard optimizers. The PVI framework automatically breaks the data set into a series of local free-energy optimization problems and the propagation of uncertainty between the data groups weights the information extracted from each. This means non-stochastic optimizers such as BFGS can now be leveraged in the large data setting. Of course, if a further stochastic approximation like Monte Carlo VI is employed for each local optimization, stochastic optimizers such as RMSProp [Tieleman and Hinton, 2012] or Adam [Kingma and Ba, 2014] might be more appropriate choices. In all cases, since the local freeenergy is equivalent in form to a global free-energy with an effective prior pe() =q(i1)()=t(i1) bi(), PVI can be implemented via trivial modification to existing code for global VI. This is a key advantage of PVI over previous local VI approaches, such as variational message passing [Winn et al., 2005, Winn and Minka, 2009, Knowles and Minka, 2011], in which bespoke and closed-form updates are needed for different likelihoods and cavity distributions. 3.3 Local Free-energy Fixed Point Updates, Natural Gradient Methods, and Mirror Descent

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

28

different likelihoods and cavity distributions. 3.3 Local Free-energy Fixed Point Updates, Natural Gradient Methods, and Mirror Descent An alternative to using off-the-shelf optimizers is to derive fixed-point update equations by zeroing the gradients of the local free-energy. These fixed-point updates have elegant properties for approximate posterior distributions that are in the exponential family. Property 4 If the prior and approximate likelihood factors are in the un-normalized exponential family tm() =tm(;m) =exp(| mT())so that the variational distribution is in the normalized exponential familyq() =exp(| qT()A(q)), then the stationary point of the local free-energydF(i)(q()) dq= 0 implies (i) bi=C1d dqEq(logp(ybij)): (8) where C:=d2A(q) dqdq=covq()[T()T|()]is the Fisher Information. Moreover, the Fisher Information can be written as C=dq

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

29

can be written as C=dq dqwhereq=Eq(T())is the mean parameter of q(). Hence, (i) bi=d dqEq(logp(ybij)): (9) For some approximate posterior distributions q(), taking derivatives of the average log-likelihood with respect to the mean parameters is analytic (e.g. Gaussian) and for some it is not (e.g. gamma). These conditions, derived in appendix A.4, can be used as fixed point equations. That is, they can be iterated possibly with damping , (i) bi= (1)(i1) bi+d dqEq(logp(ybij)): (10) These iterations, which form an inner-loop in PVI, are themselves not guaranteed to converge (there is no Lyapunov function in general and so, for example, the local free-energy will not reduce at every step). The fixed point updates are the natural gradients of the local free-energy and the damped versions are natural gradient ascent [Sato, 2001, Hoffman et al., 2013]. The natural gradients could also be used in

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

30

are natural gradient ascent [Sato, 2001, Hoffman et al., 2013]. The natural gradients could also be used in other optimization schemes [Hensman et al., 2012, Salimbeni et al., 2018]. The damped updates are also 9 equivalent to performing mirror-descent [Raskutti and Mukherjee, 2015, Khan and Lin, 2018], a general form of proximal algorithm [Parikh and Boyd, 2014] that can be interpreted as trust-region methods. For more details about the relationship between these methods, see appendix A.7. Additionally, while natural gradients or fixed-point updates have been shown to be effective in the batch global VI settings [see e.g. Honkela et al., 2010], we present some result in appendix E.6 showing adaptive first-order methods employing flat gradients such as Adam [Kingma and Ba, 2014] performs as well as natural gradient methods, when stochastic mini-batch approximations are used. For these types of updates there is an interesting relationship between PVI and global (batch) VI: Property 5 PVI methods employing parallel updates result in identical dynamics for q()given by the following equation, regardless of the partition of the data employed (i)

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

31

Property 5 PVI methods employing parallel updates result in identical dynamics for q()given by the following equation, regardless of the partition of the data employed (i) q=0+d dq(i1)Eq(logp(yj)) =0+NX n=1d dq(i1)Eq(i1)(logp(ynj)): (11) See A.5 for the proof. If parallel fixed-point updates are desired, then it is more memory efficient to employ batch VI M= 1, since then only one global set of natural parameters needs to be retained. However, as previously discussed, using M= 1gives up opportunities for online learning and distributed computation (e.g. asynchronous updates). 3.4 Stochastic mini-batch approximation There are two distinct ways to apply stochastic approximations within the PVI scheme. 3.4.1 Stochastic Approximation within the Local Free-Energy The first form of stochastic approximation leverages the fact that each local free-energy decomposes into a sum over data points and can, therefore, be approximated by sampling mini-batches within each

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

32

into a sum over data points and can, therefore, be approximated by sampling mini-batches within each data group ym. In the case where each partition includes a large number of data points, this leads to algorithms that converge more quickly than the batch variants – since a reasonable update for the approximate posterior can often be determined from just a few data points – and this faster convergence opens the door to processing larger data sets. Mini-batch approximation can be employed in the general PVI case, but for simplicity we consider the global VI case here M= 1. If simplified fixed point updates are used for optimization, then samplingLmini-batches of data from the data distribution yliidpdata(y)yields the following stochastic approximation to the damped updates,2 (i) q= (1)(i1) q + 0+Ld dqEq(logp(ylj)) ; (12) =(i1) q +0d dqEq(logp(ylj))(i1) like=L : (13)

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

33

q +0d dqEq(logp(ylj))(i1) like=L : (13) Here the first form of the update is stochastic natural gradient ascent and the second form reveals the implied deletion step where (i1) like=L= ((i1) q(i1) 0)=Lis the contribution a mini-batch likelihood makes to the posterior natural parameters on average. The rescaled learning rate is 0=L. These two forms reveals that the mini-batch stochastic natural gradient update resembles an EP update step. See appendix A.6 for full details. 2We have used a distinct notation for a mini-batch ( yl) and a data group ( ym) since the former will be selected iid from the data set and will vary at each epoch, whilst the latter need not be determined in this way and is fixed across epochs. 10 3.4.2 Stochastic Scheduling of Updates Between Local Free-Energies The second form of stochastic approximation is to randomize the update schedule. For example, using M=Nand randomly selecting subsets of data to update in parallel. This can be memory intensive,

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

34

M=Nand randomly selecting subsets of data to update in parallel. This can be memory intensive, requiringNlocal natural parameters to be stored. A more memory efficient approach is to fix the mini-batches across epochs and to visit the data groups ymin a random order [Khan and Lin, 2018]. For the simplified fixed point updates, this yields (i) m= (1)(i1) m +d dq(i1)Eq(i1)(logp(ymj)): (14) This approach results in a subtly different update to qthat retains a specific approximation to the likelihood of each data partition, rather than a single global approximation (i) q=(i1) qd dqEq(logp(ymj))(i1) m : (15) If the first approach in eq. (14) employs learning rates that obey the Robins Munro conditions, the fixed points will be identical to the second approach in eq. (15) and they will correspond to optima of the global free-energy.

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

35

fixed points will be identical to the second approach in eq. (15) and they will correspond to optima of the global free-energy. 3.4.3 Comparing and Contrasting Stochastic Approaches There are pros and cons to both approaches. The first approach in section 3.4.1 has a memory footprint Ltimes smaller than the second approach in section 3.4.2 and can converge more quickly. For example, on the first pass through the data, it effectively allows approximate likelihoods for as of yet unseen data to be updated based on those for the data seen so far, which means that larger learning rates can be used 0> . The second approach is required for continual learning, asynchronous updates, and client-side processing where the assumption that each mini-batch is iid (and a single gradient step is performed on each) is typically incorrect. The second approach also tends to produce less noisy learning curves, with stochasticity only entering via the schedule and not as an approximation to the local free-energy and the gradients thereof. These approaches could also be combined, with stochastic scheduling selecting the local free-energy to update next and mini-batch updates employed for each local free-energy optimization. See appendix A.6 for a full discussion.

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

36

to update next and mini-batch updates employed for each local free-energy optimization. See appendix A.6 for a full discussion. 4 Unification of Previous Work The local VI framework described above unifies a large number of existing approaches. These methods include global VI (section 4.1), local VI (section 4.2), online VI (section 4.3) and a number of methods based on power EP (sections 4.4 to 4.7). A schematic showing the relationships between these methods at a high level is shown in fig. 2. The literature has been organized into in fig. 3 and table 1. 4.1 Global VI Fixed Point Methods There has been a long history of applying the fixed point updates for global VI (PVI where M= 1). Sato [2001] derived them for conjugate exponential family models, showing they recover the closed form updates for q(), and noting that damped fixed point updates are equivalent to natural gradient ascent with unit step size ( = 1). Sato’s insight was subsequently built upon by several authors. Honkela et al. [2010] considered non-conjugate models, employed a Gaussian variational distribution

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

37

Honkela et al. [2010] considered non-conjugate models, employed a Gaussian variational distribution and used natural gradients to update the variational distribution’s mean. Hensman et al. [2012] and 11 Figure 2: Variational inference schemes encompassed by the PVI framework. Hoffman et al. [2013] applied the insight to conjugate models when optimizing collapsed variational free-energies and deriving stochastic natural gradient descent, respectively. Salimans and Knowles [2013] apply the fixed points to non-conjugate models where the expectations over qare intractable and use Monte Carlo to approximate them, but they explicitly calculate the Fisher information matrix, which is unnecessary for exponential family q. Sheth and Khardon [2016a] and Sheth et al. [2015] treat non-conjugate models with Gaussian latent variables, employ the cancellation of the Fisher information, and analyze convergence properties. Sheth and Khardon [2016b] further extend this to two level-models through Monte Carlo essentially applying the Fisher information cancellation to Salimans and Knowles [2013], but they were unaware of this prior work. 4.2 Fully Local VI Fixed Point Methods There has also been a long history of applying the fixed point updates in the fully local VI setting

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

38

4.2 Fully Local VI Fixed Point Methods There has also been a long history of applying the fixed point updates in the fully local VI setting (whereM=N). Knowles and Minka [2011] derive them for non-conjugate variational message passing, but explicitly calculate the Fisher information matrix (except in a case where qwas univariate Gaussian case where they do employ the cancellation). Wand [2014] simplified VMP by applying the Fisher information cancellation to the case where q()is multivariate Gaussian. Khan and Lin [2018] also extend VMP to employ MC approximation and the Fisher information cancellation. They were unaware of Wand [2014], but extend this work by treating a wider range of approximate posteriors and models, stochastic updates, and a principled approach to damping. The work is closely related to Salimans and Knowles [2013] and Sheth and Khardon [2016b], since although these papers use fixed-point updates for global VI, they show that these decompose over data points and thereby derive mini-batch updates that closely resemble fixed-point local VI. This is a result of property 5. 4.3 Online, Streaming, Incremental and Continual VI as a single pass of PVI

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

39

4.3 Online, Streaming, Incremental and Continual VI as a single pass of PVI If PVI makes a single pass through the data, the approximate likelihoods do not need to be explicitly computed or stored as data-groups are not revisited. In this case PVI reduces to initializing the approximate posterior to be the prior, q(0)() =p()and then optimizing a sequence of local freeenergies q(i)():= argmax q()2QZ dq() logq(i1)()p(ybij) q(): These have the form of standard variational inference with the prior replaced by the previous variational distribution q(i1)(). This idea – combining the likelihood from a new batch of data with the previous 12 Algorithm 2 One step of the PEP algorithm at the i-th iteration, for the bi-th data partition Compute the tilted distribution: ^p(i) () =q(i1)() p(ybij) t(i1) bi() Moment match: q() = proj(^p(i)

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

40

p(ybij) t(i1) bi() Moment match: q() = proj(^p(i) ())such that Eq()(T()) =E^p(i) ()(T()) Update the posterior distribution with damping :q(i)() = q(i1)()
1=(q())= Update the approximate likelihood: t(i) bi() =q(i)() q(i1)()t(i1) bi() Algorithm 3 One step of the PEP algorithm, as in algorithm 2, but with alpha divergence minimization Compute the tilted distribution: ^p(i)() =q(i1)()p(ybij) t(i1) bi() Find the posterior distribution: q(i)():= argminq()2QD[^p(i)()jjq()] Update the approximate likelihood: t(i) bi() =q(i)() q(i1)()t(i1) bi()

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

41

bi() =q(i)() q(i1)()t(i1) bi() approximate posterior and projecting back to a new approximate posterior – underpins online variational inference [Ghahramani and Attias, 2000, Sato, 2001], streaming variational inference [Broderick et al., 2013, Bui et al., 2017b], and variational continual learning [Nguyen et al., 2018]. Early work on online VI used conjugate models and analytic updates [Ghahramani and Attias, 2000, Sato, 2001, Broderick et al., 2013, Bui et al., 2017b], this was followed by off-the-shelf optimization approaches for non-conjugate models [Bui et al., 2017b] and further extended to leverage MC approximations of the local-free energy [Nguyen et al., 2018]. Recently Zeno et al. [2018] use the variational continual learning framework of Nguyen et al. [2018], but employ fixed-point updates instead. 4.4 Power EP as a Fully Local VI Fixed Point Method There is also an important relationship between PVI methods employing fixed point updates and power

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

42

4.4 Power EP as a Fully Local VI Fixed Point Method There is also an important relationship between PVI methods employing fixed point updates and power expectation propagation [Minka, 2004]. Property 6 below states that the local VI fixed point equations are recovered from the Power EP algorithm as !0. Property 6 The damped fixed point equations are precisely those returned by the PEP algorithm, shown in algorithm 2, in the limit that !0. Although we suspect Knowles and Minka [2011] knew of this relationship, and it is well known that Power EP has the same fixed points as VI in this case, it does not appear to be widely known that variationally limited Power EP yields exactly the same algorithm as fixed point local VI. See A.8 for the proof. 4.5 Alpha-divergence EP as a Local VI Method with Off-the-shelf Optimization PVI is intimately related to alpha-divergence EP. If PVI’s KL divergence is replaced by an alpha divergence D[p()jjq()] =1 (1)R

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

43

divergence D[p()jjq()] =1 (1)R p() + (1)q()p()q()1 dwe recover the alphadivergence formulation of the power-EP algorithm [Minka, 2004] which encompasses the current case as !0and EP when !1[Minka, 2001]. The updates using this formulation are shown in algorithm 3. The alpha divergence is typically very difficult to compute once more than one non-Gaussian likelihood is included in a data group ym, meaning that for general alpha it would be appropriate to set M=N. The variational KL is the exception as it decomposes over data points. 13 Algorithm 4 One step of the SPEP algorithm at the i-th iteration, for the bi-th data partition Compute the tilted distribution: ^p(i) () =q(i1)()p(ybij) t(i1)() Moment match: q() = proj(^p(i) ())such that Eq()(T()) =E^p(i)

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

44

())such that Eq()(T()) =E^p(i) ()(T()) Update the posterior distribution with damping :q(i)() = q(i1)()
1N=(q())N= Update the approximate likelihood: t(i)= q(i)() p()1=N 4.6 Stochastic Power EP as a Stochastic Global VI Fixed Point Method The stochastic power EP algorithm [Li et al., 2015] reduces the memory overhead of EP by maintaining a single likelihood approximation that approximates the average effect a likelihood has on the posterior q() =p()t()M. Taking the variational limit of this algorithm, !0, we recover global VI M= 1 with damped simplified fixed-point updates that employ a stochastic (mini-batch) approximation [Hoffman et al., 2013]. Property 7 The mini-batch fixed point equations are precisely those returned by the SPEP algorithm, shown in algorithm 4 in the limit that !0.

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127

1811.11206

45

Property 7 The mini-batch fixed point equations are precisely those returned by the SPEP algorithm, shown in algorithm 4 in the limit that !0. In this way the relationship between EP and SEP is the same as the relationship between fixed point PVI and fixed point mini-batch global VI (see section 3.4 where the two approaches differ by removing either an average natural parameter or a specific one). Similarly, if we altered PVI to maintain a single average likelihood approximation, as SEP does, we would recover mini-batch global VI. 4.7 Distributed (Power) EP Methods The convergent distributed Power EP approach of Hasenclever et al. [2017] recovers a version of PVI as !0with convergence guarantees. The PVI approach is also similar in spirit to Gelman et al. [2014], Hasenclever et al. [2017] who use EP to split up data sets into small parts that are amenable to MCMC. Here we are using PVI to split up data sets so that they are amenable for optimization. 5 ImprovingFederatedLearningfor BayesianNeuralNetworksUsing PVI Having connected the previous literature using the unifying framework based on PVI, we will discuss

1811.11206

Partitioned Variational Inference: A unified framework encompassing federated and continual learning

Variational inference (VI) has become the method of choice for fitting many modern probabilistic models. However, practitioners are faced with a fragmented literature that offers a bewildering array of algorithmic options. First, the variational family. Second, the granularity of the updates e.g. whether the updates are local to each data point and employ message passing or global. Third, the method of optimization (bespoke or blackbox, closed-form or stochastic updates, etc.). This paper presents a new framework, termed Partitioned Variational Inference (PVI), that explicitly acknowledges these algorithmic dimensions of VI, unifies disparate literature, and provides guidance on usage. Crucially, the proposed PVI framework allows us to identify new ways of performing VI that are ideally suited to challenging learning scenarios including federated learning (where distributed computing is leveraged to process non-centralized data) and continual learning (where new data and tasks arrive over time and must be accommodated quickly). We showcase these new capabilities by developing communication-efficient federated training of Bayesian neural networks and continual learning for Gaussian process models with private pseudo-points. The new methods significantly outperform the state-of-the-art, whilst being almost as straightforward to implement as standard VI.

http://arxiv.org/pdf/1811.11206

stat.ML

20181127

20181127