I-JEPA is a method for self-supervised learning. At a high level, I-JEPA predicts a representation of one part of an image from representations of other parts of the same image. Notably, this approach learns semantic image features:

1. Do not rely on pre-specified invariants for hand-crafted data transformations, which tend to be biased towards specific downstream tasks,
2. And without filling the model with pixel-level details, this often results in learning less semantically meaningful representations.

In contrast to generative methods with pixel decoders, I-JEPA has predictors that make predictions in the latent space. The predictor in I-JEPA can be viewed as a primitive (and constrained) world model capable of simulating spatial uncertainty in static images from partially observable context.This world model is semantic in that it predicts high-level information about unseen regions in the image, rather than pixel-level detail.

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

I-JEPA pretraining is also computationally efficient. It does not involve any overhead associated with applying more computationally intensive data augmentation to generate multiple views. The target encoder only needs to process one view of the image, and the context encoder only needs to process the context block. Empirically, I-JEPA learns powerful off-the-shelf semantic representations without manual view augmentation.

pre-trained model