Research map / Label-efficient and robust learning / Self-supervised and contrastive
Local plasticity rules can learn deep representations using self-supervised contrastive predictions
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 |
|---|---|---|---|
| Block gradients between layers (ablation) | Hinge Loss CPC on STL-10 · STL-10 | linear classification accuracy: almost 5% | Hurt |
| Increase the number of gradient-isolated modules (ablation) | Hinge Loss CPC on STL-10 · STL-10 | linear classification accuracy: Performance drops monotonously but not catastrophically | Hurt |
| Use hinge loss instead of CPC cross-entropy (ablation) | CPC on STL-10 · STL-10 | linear classification accuracy: less than 1% | No clear effect |
| Use temporally local inputs instead of synchronous negative samples (ablation) | Hinge Loss CPC on STL-10 · STL-10 | linear classification accuracy: 1.2% | Hurt |
| Use separate retrodiction weights instead of the prediction-weight transpose (ablation) | Hinge Loss CPC on STL-10 · STL-10 | linear classification accuracy: 0.1% | No clear effect |
| Remove the retrodiction-weight update (ablation) | Hinge Loss CPC on STL-10 · STL-10 | linear classification accuracy: 2% | Hurt |
| Replace backpropagation through time with e-prop (ablation) | CLAPP-s on LibriSpeech · LibriSpeech | phone classification accuracy: 3.1 % | Hurt |
| Remove the GRU layer (ablation) | CLAPP-s on LibriSpeech · LibriSpeech | phone classification accuracy: 9.3 % | Hurt |
| Use layer-wise rather than end-to-end training (ablation) | CLAPP-s on LibriSpeech · LibriSpeech | phone classification accuracy: 1.1% | Hurt |
| Use layer-wise rather than end-to-end training (ablation) | Layer-wise GIM on LibriSpeech · LibriSpeech | phone classification accuracy: 0.4% | Hurt |
| Block gradients between layers (ablation) | CLAPP-s on STL-10 · STL-10 | linear classification accuracy: 80.3 -> 75.0 | Hurt |
Closest papers by meaning
- Characterizing emergent representations in a space of candidate learning rules for deep networks (NeurIPS 2020 · similarity 0.88)
- Dendritic Localized Learning: Toward Biologically Plausible Algorithm (ICML 2025 · similarity 0.87)
- Contrastive-Equivariant Self-Supervised Learning Improves Alignment with Primate Visual Area IT (NeurIPS 2024 · similarity 0.87)
- Learning the Plasticity: Plasticity-Driven Learning Framework in Spiking Neural Networks (NeurIPS 2025 · similarity 0.86)
- Self-Attention-Based Contextual Modulation Improves Neural System Identification (ICLR 2025 · similarity 0.86)
- Flexible Context-Driven Sensory Processing in Dynamical Vision Models (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.