jepa-lab

Small, high-leverage reproductions for JEPA-style representation learning and recent Yann LeCun world-model work.

Lanes

1. 01-slow-features - toy failure-mode reproduction for static distractors and slow-feature bias.
2. 02-mini-ijepa - minimal image JEPA with masked-region target embeddings.
3. 03-lejepa-sigreg - stability and anti-collapse experiments using SIGReg-style regularization.
4. 04-eb-jepa - compact image, video, and planning reproductions from the EB-JEPA library.
5. 05-leworldmodel - latent world-model and planning experiments on small control tasks.
6. 06-vjepa-probes - checkpoint-based probing and reduced-scale video JEPA experiments.

Layout

configs/       experiment configs
data/          local datasets, ignored by git
notes/         roadmap, paper notes, and experiment log
papers/        one directory per reproduction lane
runs/          local outputs, checkpoints, and logs, ignored by git
scripts/       one-off commands and analysis helpers
src/jepa_lab/  shared utilities

First Milestones

1. Reproduce the slow-feature toy result with a tiny synthetic dataset.
2. Build a mini I-JEPA training loop on CIFAR-10 or Tiny-ImageNet.
3. Add SIGReg as a narrow ablation against the mini I-JEPA baseline.