I-JEPA

Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture

Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, Nicolas Ballas · CVPR 2023

Read paper · arXiv:2301.08243

Predict teacher representations of masked image blocks from context.

What you will implement

Target-block gathering, prediction loss, and EMA teacher updates.

• 57. I-JEPA Masked Latent Targets and EMA

  Extract frozen target blocks, train latent predictions and maintain momentum target parameters.

  Implement target layer normalization, masked extraction, smooth L1 and EMA from the author code; target features and masks are supplied, without a ViT or mask sampler.