Research map / Label-efficient and robust learning / Self-supervised and contrastive
DSeq-JEPA: Discriminative Sequential Joint-Embedding Predictive Architecture
Xiangteng He, Shunsuke SAKAI, Shivam Chandhok, Sara Beery, Kun yuan, Nicolas Padoy, Tatsuhito Hasegawa, Leonid Sigal
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 |
|---|---|---|---|
| add discriminative sequential prediction to I-JEPA | I-JEPA · ImageNet-1K linear probe (ViT-B/16) | top-1 accuracy: 72.4 -> 73.5 | Improved |
| add discriminative sequential prediction to I-JEPA | I-JEPA · ImageNet-1K full fine-tune (ViT-B/16) | top-1 accuracy: 83.5 -> 84.0 | Improved |
| add discriminative sequential prediction to I-JEPA | I-JEPA · ImageNet-1K linear probe (ViT-H/16) | top-1 accuracy: 81.1 -> 82.4 | Improved |
| add discriminative sequential prediction to C-JEPA | C-JEPA · ImageNet-1K linear probe (ViT-B/16) | top-1 accuracy: 73.5 -> 73.8 | Improved |
| add discriminative sequential prediction to I-JEPA | I-JEPA · CUB-200-2011 and Cars (ViT-B/16) | top-1 accuracy: +0.9 on CUB, +1.4 on Cars | Improved |
| add discriminative region selection and sequential prediction together vs either alone (ablation) | DSeq-JEPA · ImageNet-1K linear probe (ViT-B/16) | top-1 accuracy: 72.4 -> 73.5 | Improved |
| add sequential prediction alone with uniform sampling (ablation) | I-JEPA · ImageNet-1K linear probe (ViT-B/16) | top-1 accuracy: 72.4 -> 72.3 | Hurt |
| add discriminative selection alone with flat prediction (ablation) | I-JEPA · ImageNet-1K linear probe (ViT-B/16) | top-1 accuracy: 72.4 -> 72.0 | Hurt |
| use discriminative order rather than random order (ablation) | DSeq-JEPA · ImageNet-1K linear probe (ViT-B/16) | top-1 accuracy: 71.7 -> 73.5 | Improved |
| use discriminative order rather than spatial order (ablation) | DSeq-JEPA · ImageNet-1K linear probe (ViT-B/16) | top-1 accuracy: 72.7 -> 73.5 | Improved |
| use forward discriminative order rather than inverse order (ablation) | DSeq-JEPA · ImageNet-1K linear probe (ViT-B/16) | top-1 accuracy: 71.3 -> 73.5 | Improved |
| predict all Top-5 regions rather than truncating at Top-3 (ablation) | DSeq-JEPA · ImageNet-1K linear probe (ViT-B/16) | top-1 accuracy: 73.0 -> 73.5 | Improved |
Closest papers by meaning
- Connecting Joint-Embedding Predictive Architecture with Contrastive Self-supervised Learning (NeurIPS 2024 · similarity 0.91)
- How JEPA Avoids Noisy Features: The Implicit Bias of Deep Linear Self Distillation Networks (NeurIPS 2024 · similarity 0.90)
- seq-JEPA: Autoregressive Predictive Learning of Invariant-Equivariant World Models (NeurIPS 2025 · similarity 0.90)
- Denoising with a Joint-Embedding Predictive Architecture (ICLR 2025 · similarity 0.88)
- LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels (NeurIPS 2026 · similarity 0.86)
- You Don’t Need Domain-Specific Data Augmentations When Scaling Self-Supervised Learning (NeurIPS 2024 · similarity 0.86)
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.
