Research map / Reinforcement learning / Policy gradient and actor-critic
Learning to solve the credit assignment problem
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 |
|---|---|---|---|
| Learn feedback weights using node-perturbation gradient estimates instead of fixed feedback weights | CNN · CIFAR10 | test accuracy: 72.4$\pm$0.2 -> 74.8$\pm$0.2 | Helped |
| Learn feedback weights using node-perturbation gradient estimates instead of fixed feedback weights | CNN · CIFAR100 | test accuracy: 47.3$\pm$0.1 -> 48.1$\pm$0.2 | Helped |
| Use node perturbation rather than backpropagation | CNN · CIFAR10 | test accuracy: 76.9$\pm$0.1 -> 74.8$\pm$0.2 | Hurt |
| Use node perturbation rather than backpropagation | CNN · CIFAR100 | test accuracy: 51.2$\pm$0.1 -> 48.1$\pm$0.2 | Hurt |
| Add noise to activations (ablation) | Backpropagation with SGD · MNIST autoencoder | mean loss: 609.8$\pm$14.4 -> 536.8$\pm$2.1 | Helped |
| Add noise to activations (ablation) | Backpropagation with ADAM · MNIST autoencoder | mean loss: 533.3$\pm$2.2 -> 522.3$\pm$0.4 | Helped |
| Add noise to activations (ablation) | Feedback alignment · MNIST autoencoder | mean loss: 759.1$\pm$3.3 -> 768.2$\pm$2.7 | No clear effect |
| Use node-perturbation estimates to train feedback weights rather than fixed feedback weights | Feedback alignment · MNIST autoencoder | mean loss: 768.2$\pm$2.7 -> 515.3$\pm$4.1 | Helped |
| Use noisy node perturbation rather than synthetic gradients supervised by true gradients (ablation) | Feedback-weight learning · MNIST autoencoder | mean loss: 521.6$\pm$2.3 -> 515.3$\pm$4.1 | Helped |
| Add noise to activations (ablation) | Backpropagation · CIFAR10 | mean accuracy: 76.9$\pm$0.1 -> 76.8$\pm$0.2 | No clear effect |
| Add noise to activations (ablation) | Direct feedback alignment · CIFAR10 | mean accuracy: 72.3$\pm$0.1 -> 72.4$\pm$0.2 | No clear effect |
| Learn feedback weights using true gradients rather than fixed feedback weights (ablation) | Synthetic gradients · CIFAR10 | mean accuracy: 72.3$\pm$0.1 -> 75.3$\pm$0.3 | Helped |
Closest papers by meaning
- Minimizing Control for Credit Assignment with Strong Feedback (ICML 2022 · similarity 0.90)
- Structural Credit Assignment in Neural Networks using Reinforcement Learning (NeurIPS 2021 · similarity 0.90)
- Forward Learning with Top-Down Feedback: Empirical and Analytical Characterization (ICLR 2024 · similarity 0.87)
- Efficient Reinforcement Learning by Discovering Neural Pathways (NeurIPS 2024 · similarity 0.87)
- Feedback control guides credit assignment in recurrent neural networks (NeurIPS 2024 · similarity 0.86)
- Measuring and Controlling Solution Degeneracy across Task-Trained Recurrent Neural Networks (NeurIPS 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.