An interactive, from-scratch implementation of LeJEPA in JAX/Equinox.
Heuristics-free self-supervised representation learning for time series with SIGReg (LeJEPA). Disentangles time-axis collapse, positional structure, and representation richness across PatchTST, TCN, and bag-of-patches encoders. Reproducible, seeded, significance-tested.
From-scratch JEPA world-model engine in PyTorch (FSDP, bf16, SIGReg/VICReg): data curation, distributed training and inference optimisation, evaluated on CAMELS cosmology against control floors with seed noise bands.
To associate your repository with the sigreg topic, visit your repo's landing page and select "manage topics."