Research map / Architectures and efficiency / RNNs and normalization
Minimizing Control for Credit Assignment with Strong Feedback
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 strong feedback with noisy neural dynamics instead of weak-feedback DFC | DFC · MNIST | test error: 15.15^{\pm 0.44} -> 2.19^{\pm 0.05} | Helped |
| Use strong feedback with noisy neural dynamics instead of weak-feedback DFC | DFC · Fashion-MNIST | test error: 16.29^{\pm 0.41} -> 12.07^{\pm 0.16} | Helped |
| Add noise during forward-weight training (ablation) | Strong-DFC · MNIST | test error: 1.98^{\pm 0.05} -> 2.19^{\pm 0.05} | Hurt |
| Add noise during forward-weight training (ablation) | Strong-DFC · Fashion-MNIST | test error: 11.36^{\pm 0.17} -> 12.07^{\pm 0.16} | Hurt |
| Use noisy rather than noiseless forward-weight training | DFC · MNIST | test error: 2.09^{\pm 0.10} -> 15.15^{\pm 0.44} | Hurt |
| Use noisy rather than noiseless forward-weight training | DFC · Fashion-MNIST | test error: 11.31^{\pm 0.14} -> 16.29^{\pm 0.41} | Hurt |
| Use ideal feedback weights rather than BP on the MNIST MLP | Strong-DFC · MNIST | test error: 1.78^{\pm 0.04}\% -> 1.98^{\pm 0.11}\% | Hurt |
| Train feedback weights concurrently using the proposed anti-Hebbian rule rather than leave them fixed at random initialization (ablation) | Strong-DFC · student-teacher regression | training loss | Helped |
| Add the delay-compensating scaling term to feedback-weight updates (ablation) | Strong-DFC · student-teacher regression | training loss | Helped |
| Pre-train feedback weights before full network training (ablation) | Strong-DFC · student-teacher regression | training stability | Helped |
| Add hidden-layer training rather than train only the output layer (ablation) | Strong-DFC · nonlinear student-teacher function approximation | training loss | Helped |
Closest papers by meaning
- Credit Assignment in Neural Networks through Deep Feedback Control (NeurIPS 2021 · similarity 0.94)
- Learning to solve the credit assignment problem (ICLR 2020 · similarity 0.90)
- Feedback control guides credit assignment in recurrent neural networks (NeurIPS 2024 · similarity 0.89)
- Forward Learning with Top-Down Feedback: Empirical and Analytical Characterization (ICLR 2024 · similarity 0.87)
- Overcoming Rank Collapse in Feedback Alignment (NeurIPS 2026 · similarity 0.85)
- Counter-Current Learning: A Biologically Plausible Dual Network Approach for Deep Learning (NeurIPS 2024 · 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.