Research map / Probabilistic and scientific modeling / Dynamical systems
Identifying Learning Rules From Neural Network Observables
Aran Nayebi, Sanjana Srivastava, Surya Ganguli, Daniel L Yamins
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 |
|---|---|---|---|
| Use aggregate statistics of weights, activations, or layer-wise activity changes across the full learning trajectory | Random Forest learning-rule classifier · Architectures, tasks, and training curricula represented in the study | learning-rule identification accuracy | Improved |
| Use a Random Forest instead of a linear SVM or Conv1D MLP | Learning-rule identification · Full trajectories and observable measures | test accuracy | Improved |
| Use activation statistics rather than weight or layer-wise activity-change statistics | Learning-rule identification with unit subsampling and measurement noise · FA vs. IA on ResNet-18, ImageNet and SimCLR | test accuracy | Improved |
| Measure trajectory points further apart in time | Learning-rule identification from subsampled trajectories · Multiple observable measures | test accuracy | Improved |
| Average observable statistics over the full learning trajectory instead of retaining the trajectory (ablation) | SVM learning-rule classifier · All observable measures | test accuracy | Hurt |
| Average observable statistics over the full learning trajectory instead of retaining the trajectory (ablation) | Random Forest learning-rule classifier · Weight observables | test accuracy | No clear effect |
| Train only on one consecutive portion of the learning trajectory rather than sampling across it (ablation) | Random Forest learning-rule classifier · Trajectory undersampling | generalization performance | Hurt |
| Use weight observables rather than other individual observable measures | Learning-rule identification with trajectory undersampling · Trajectory subsampling | test accuracy | Improved |
| Use activation observables rather than weight or layer-wise activity-change observables | Learning-rule identification · Unit undersampling and measurement noise | test accuracy | Improved |
| Generalize from supervised cross-entropy training to self-supervised SimCLR (ablation) | Random Forest learning-rule classifier · Held-out task | test accuracy | Hurt |
| Use a linear classifier rather than a non-linear classifier (ablation) | Learning-rule identification from individual observables · Consecutive portions of the learning trajectory | test accuracy | Hurt |
| Identify learning rules from activation trajectories rather than temporally averaged post-synaptic activities | Learning-rule identification · Multiple learning-rule classes | separability | Improved |
Closest papers by meaning
- Model Based Inference of Synaptic Plasticity Rules (NeurIPS 2024 · similarity 0.90)
- Flexible inference for animal learning rules using neural networks (NeurIPS 2025 · similarity 0.90)
- Dynamics of Supervised and Reinforcement Learning in the Non-Linear Perceptron (NeurIPS 2024 · similarity 0.88)
- Discovering plasticity rules that organize and maintain neural circuits (NeurIPS 2024 · similarity 0.88)
- Learning the Plasticity: Plasticity-Driven Learning Framework in Spiking Neural Networks (NeurIPS 2025 · similarity 0.88)
- How connectivity structure shapes rich and lazy learning in neural circuits (ICLR 2024 · similarity 0.87)
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.
