Research map / Label-efficient and robust learning / Self-supervised and contrastive
Self-Supervised Learning From Images With a Joint-Embedding Predictive Architecture
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use I-JEPA representation prediction instead of pixel reconstruction (ablation) | I-JEPA · ImageNet-1K 1% labels | Top-1: 40.7 -> 66.9 | Helped |
| Use multi-block masking instead of rasterized quadrant masking (ablation) | I-JEPA · ImageNet-1K 1% labels | Top-1: 15.5 -> 54.2 | Helped |
| Use multi-block masking instead of single-block masking (ablation) | I-JEPA · ImageNet-1K 1% labels | Top-1: 20.2 -> 54.2 | Helped |
| Use multi-block masking instead of random-patch masking (ablation) | I-JEPA · ImageNet-1K 1% labels | Top-1: 17.6 -> 54.2 | Helped |
| Predict target blocks with scale (0.15, 0.2) rather than the tested smaller or larger scales (ablation) | I-JEPA multi-block masking · ImageNet-1K 1% labels | Top-1: 19.2, 39.2, 42.4, 54.2, 38.9, 33.6 | Mixed |
| Use a larger context block, with scale (0.85, 1.0) (ablation) | I-JEPA multi-block masking · ImageNet-1K 1% labels | Top-1: 31.2, 47.1, 49.3, 54.2 | Helped |
| Increase the number of target blocks (ablation) | I-JEPA multi-block masking · ImageNet-1K 1% labels | Top-1: 9.0, 22.0, 48.5, 54.2 | Helped |
| Mask the target-encoder output rather than its input (ablation) | I-JEPA · ImageNet-1K 1% labels | Top-1: 56.1 -> 67.3 | Helped |
| Increase predictor depth from 6 to 12 (ablation) | I-JEPA · ImageNet-1K 1% labels | Top-1: 64.0 -> 66.9 | Helped |
| Use fixed weight decay 0.05 instead of increasing it from 0.04 to 0.4 (ablation) | I-JEPA · ImageNet-1K | Top-1: 69.4 -> 70.7; 77.8 -> 76.4 | Mixed |
| Use a predictor width bottleneck of 384 rather than width 1024 (ablation) | I-JEPA · ImageNet-1K 1% labels, fine-tuning | Top-1: 68.4 -> 70.7 | Helped |
| Pretrain on the larger IN22k dataset instead of IN1k (ablation) | I-JEPA ViT-H/14 · CIFAR100 linear probe | accuracy: 87.5 -> 89.5 | Helped |
Closest papers by meaning
- How JEPA Avoids Noisy Features: The Implicit Bias of Deep Linear Self Distillation Networks (NeurIPS 2024 · similarity 0.91)
- Connecting Joint-Embedding Predictive Architecture with Contrastive Self-supervised Learning (NeurIPS 2024 · similarity 0.91)
- You Don’t Need Domain-Specific Data Augmentations When Scaling Self-Supervised Learning (NeurIPS 2024 · similarity 0.88)
- Denoising with a Joint-Embedding Predictive Architecture (ICLR 2025 · similarity 0.88)
- seq-JEPA: Autoregressive Predictive Learning of Invariant-Equivariant World Models (NeurIPS 2025 · similarity 0.87)
- LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels (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.