jmtzt/ijepa_huge_patch14_22k

I-JEPA Model (Huge, fine-tuned on IN22K)

I-JEPA is a method for self-supervised learning. At a high level, I-JEPA predicts the representations of part of an image from the representations of other parts of the same image:

1. without relying on pre-specified invariances to hand-crafted data transformations, which tend to be biased for particular downstream tasks,
2. and without having the model fill in pixel-level details, which tend to result in learning less semantically meaningful representations.

How does it work?

As opposed to generative methods that have a pixel decoder, I-JEPA has a predictor that makes predictions in latent space. The predictor in I-JEPA can be seen as a primitive (and restricted) world-model that is able to model spatial uncertainty in a static image from a partially observable context. This world model is semantic in the sense that it predicts high level information about unseen regions in the image, rather than pixel-level details.

We trained a stochastic decoder that maps the I-JEPA predicted representations back in pixel space as sketches. The model correctly captures positional uncertainty and produces high-level object parts with the correct pose (e.g., dog’s head, wolf’s front legs).

Intended uses & limitations

I-JEPA can be used for image classification or feature extraction. This checkpoint in specific is intended for Feature Extraction.

BibTeX entry and citation info

If you use I-JEPA or this code in your work, please cite:

@article{assran2023self,
  title={Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture},
  author={Assran, Mahmoud and Duval, Quentin and Misra, Ishan and Bojanowski, Piotr and Vincent, Pascal and Rabbat, Michael and LeCun, Yann and Ballas, Nicolas},
  journal={arXiv preprint arXiv:2301.08243},
  year={2023}
}