Research map / Label-efficient and robust learning / Self-supervised and contrastive
Rethinking JEPA: Compute‑Efficient Video Self-Supervised Learning with Frozen Teachers
Xianhang Li, Chen Huang, Chun-Liang Li, Eran Malach, Joshua M. Susskind, Vimal Thilak, Etai Littwin
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace EMA teacher with frozen pixel-reconstruction teacher (SALT) (ablation) | V-JEPA 2 ViT-L on V-3.6M · 6-benchmark average | avg top-1 accuracy: 79.6 -> 81.9 | Improved |
| replace EMA teacher with frozen teacher (SALT) | V-JEPA 2 ViT-L on V-3.6M · SSv2 (16x2x3) | top-1 accuracy: 68.2 -> 74.9 | Improved |
| train student with stage 2 latent prediction after teacher (ablation) | SALT ViT-L · 6-benchmark average | avg top-1 accuracy: 77.3 -> 81.9 | Improved |
| use ViT-L teacher instead of larger ViT-H/ViT-G teacher (ablation) | SALT ViT-L student · avg of K400/SSv2/IN1k/COIN | avg top-1 accuracy: 77.4 vs 77.3 and 77.6 | No clear effect |
| use ViT-L teacher instead of ViT-H teacher (ablation) | SALT ViT-G student · avg of K400/SSv2/IN1k/COIN | avg top-1 accuracy: 78.7 -> 79.0 | Improved |
| teacher pretrained on ImageNet-1k instead of video data (ablation) | SALT student · avg of K400/SSv2/IN1k/COIN | avg top-1 accuracy: 77.4 -> 74.9 | Hurt |
| teacher pretrained on SSv2 only instead of V-3.6M (ablation) | SALT student · avg of K400/SSv2/IN1k/COIN | avg top-1 accuracy: 77.4 -> 75.5 | Hurt |
| teacher pretrained on K710 only instead of V-3.6M (ablation) | SALT student · avg of K400/SSv2/IN1k/COIN | avg top-1 accuracy: 77.4 -> 77.4 | No clear effect |
| multi-block teacher masking instead of 2x random tube (ablation) | SALT student · avg of K400/SSv2/IN1k/COIN | avg top-1 accuracy: 76.9 -> 77.4 | Improved |
| multi-block teacher masking instead of causal masking (ablation) | SALT student · avg of K400/SSv2/IN1k/COIN | avg top-1 accuracy: 73.7 -> 77.4 | Improved |
| shorter teacher training (40k of 240k steps) instead of 80k (ablation) | SALT ViT-L · avg top-1 accuracy | avg top-1 accuracy: 77.9 -> 78.2 | Improved |
| use SALT student loss as model selection signal | SALT student · SSv2 | R2 of loss vs accuracy: R2 = 0.951 | Improved |
Closest papers by meaning
- How JEPA Avoids Noisy Features: The Implicit Bias of Deep Linear Self Distillation Networks (NeurIPS 2024 · similarity 0.88)
- Connecting Joint-Embedding Predictive Architecture with Contrastive Self-supervised Learning (NeurIPS 2024 · similarity 0.87)
- seq-JEPA: Autoregressive Predictive Learning of Invariant-Equivariant World Models (NeurIPS 2025 · similarity 0.86)
- Extending Video Masked Autoencoders to 128 frames (NeurIPS 2024 · similarity 0.85)
- Denoising with a Joint-Embedding Predictive Architecture (ICLR 2025 · similarity 0.85)
- PEIRA: Learning Predictive Encoders through Inter-View Regressor Alignment (NeurIPS 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
