Research map / Architectures and efficiency / RNNs and normalization
Credit Assignment in Neural Networks through Deep Feedback Control
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 DFC with learned feedback weights | DFC · MNIST-autoencoder | test loss: 33.37^{\pm0.60} \cdot 10^{-2} -> 11.28^{\pm 0.18} \cdot 10^{-2} | Helped |
| Fix feedback weights at initialization instead of learning them (ablation) | DFC · MNIST-autoencoder | test loss: 11.28^{\pm 0.18} \cdot 10^{-2} -> 33.37^{\pm0.60} \cdot 10^{-2} | Hurt |
| Update forward weights only after reaching steady state (ablation) | DFC · MNIST-autoencoder | test loss: 11.28^{\pm 0.18} \cdot 10^{-2} -> 11.21^{\pm 0.04} \cdot 10^{-2} | No clear effect |
| Compute the linearized steady state analytically (ablation) | DFC · MNIST-autoencoder | test loss: 11.28^{\pm 0.18} \cdot 10^{-2} -> 11.27^{\pm 0.09} \cdot 10^{-2} | No clear effect |
| Use DFC rather than BP | DFC · MNIST (train loss) | train loss: 1.53^{\pm 0.19}\cdot 10^{-7} -> 7.61^{\pm 0.65}\cdot 10^{-8} | Helped |
| Set lambda to 0.5 (ablation) | DFC · MNIST | test error: 2.30^{\pm 0.06} -> 2.34^{\pm 0.15} | No clear effect |
| Vary lambda from 1e^{-3} to 0.5 (ablation) | DFC-SS · MNIST | test error: 2.26^{\pm 0.11} -> 2.31^{\pm 0.15} | No clear effect |
| Vary alpha from 1e^{-4} to 1e^{-1} (ablation) | DFC · MNIST | test error: 2.28^{\pm 0.06} -> 2.34^{\pm 0.16} | No clear effect |
| Fix feedback weights rather than learning them (ablation) | DFC · Fashion-MNIST | test error: 11.17^{\pm 0.27}\% -> 11.62^{\pm 0.30}\% | Hurt |
| Use DFC instead of DFA | DFC · MNIST-autoencoder | test loss: 29.95^{\pm 0.36} \cdot 10^{-2} -> 11.28^{\pm 0.18} \cdot 10^{-2} | Helped |
Closest papers by meaning
- Minimizing Control for Credit Assignment with Strong Feedback (ICML 2022 · similarity 0.94)
- Feedback control guides credit assignment in recurrent neural networks (NeurIPS 2024 · similarity 0.91)
- Biological credit assignment through dynamic inversion of feedforward networks (NeurIPS 2020 · similarity 0.91)
- Credit Assignment via Neural Manifold Noise Correlation (ICML 2026 · similarity 0.88)
- Characterizing control between interacting subsystems with deep Jacobian estimation (NeurIPS 2025 · similarity 0.86)
- Improving Deep Learning Speed and Performance Through Synaptic Neural Balance (AAAI 2025 · 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.