Research map / Label-efficient and robust learning / Self-supervised and contrastive
Can local learning match self-supervised backpropagation?
Wu S. Zihan, Ariane Delrocq, Wulfram Gerstner, Guillaume Bellec
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 2D spatial dependence to feedback B (ablation) | CLAPP++ · CIFAR-10 linear probe | accuracy: 73.21 -> 80.51 | Improved |
| add 2D spatial dependence to feedback B (ablation) | CLAPP++ · STL-10 linear probe | accuracy: 75.10 -> 78.66 | Improved |
| add 2D spatial dependence to feedback B (ablation) | CLAPP++ · Tiny-ImageNet linear probe | accuracy: 28.18 -> 36.63 | Improved |
| add 2D spatial dependence to feedback B (ablation) | CLAPP++ · ImageNet linear probe | accuracy: 38.31 -> 42.55 | Improved |
| add direct feedback from last layer (ablation) | CLAPP++ · STL-10 linear probe | accuracy: 78.66 -> 79.38 | Improved |
| add direct feedback from last layer (ablation) | CLAPP++ · ImageNet linear probe | accuracy: 42.55 -> 44.16 | Improved |
| optimise sum of direct-feedback and local losses (both) (ablation) | CLAPP++ · CIFAR-10 linear probe | accuracy: 80.51 -> 81.18 | Improved |
| optimise sum of direct-feedback and local losses (both) (ablation) | CLAPP++ · ImageNet linear probe | accuracy: 42.55 -> 42.49 | No clear effect |
| use local-SSL instead of end-to-end BP with same loss | CLAPP++ · ImageNet linear probe | accuracy: 48.52 -> 42.55 | Hurt |
| use local-SSL instead of BP | CLAPP++ · VRAM, STL-10 (6-layer VGG) | peak VRAM (GB): 5.47 -> 3.38 | Improved |
| use local-SSL instead of BP | CLAPP++ · VRAM, ImageNet (8-layer VGG) | peak VRAM (GB): 11.10 -> 4.60 | Improved |
| add spatial dependence to BP-SSL (ablation) | BP-CLAPP++ · benchmarks | accuracy | No clear effect |
Closest papers by meaning
- Effective Self-supervised Pre-training on Low-compute Networks without Distillation (ICLR 2023 · similarity 0.87)
- A Closer Look at Benchmarking Self-supervised Pre-training with Image Classification (IJCV 2025 · similarity 0.86)
- Dynamic Data Selection for Efficient SSL via Coarse-to-Fine Refinement (ECCV 2024 · similarity 0.85)
- Towards Interpretable Deep Local Learning with Successive Gradient Reconciliation (ICML 2024 · similarity 0.85)
- On the Importance of Embedding Norms in Self-Supervised Learning (ICML 2025 · similarity 0.83)
- Localizing Memorization in SSL Vision Encoders (NeurIPS 2024 · similarity 0.83)
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.
