Action-conditioned JEPA world model for two-room navigation in 89K parameters. VICReg plus an inverse-dynamics auxiliary loss for collapse resistance; 4.40 MSE on normal probes.
-
Updated
Sep 20, 2026 - Jupyter Notebook
Action-conditioned JEPA world model for two-room navigation in 89K parameters. VICReg plus an inverse-dynamics auxiliary loss for collapse resistance; 4.40 MSE on normal probes.
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.
[NeurIPS 2024] Elliptical Attention
LLaVA/CLIP representation probes and a Kaiming-initialized non-exclusive EBR reproduction
To associate your repository with the representation-collapse topic, visit your repo's landing page and select "manage topics."