Research map / Architectures and efficiency / RNNs and normalization
Understanding and Controlling Memory in Recurrent Neural Networks
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 delay curriculum instead of vocabulary curriculum | GRU and LSTM delayed classification RNNs · Long-delay extrapolation on MNIST and CIFAR-10 | retrieval accuracy | Helped |
| Reduce the speeds of class-associated slow points | Delayed classification RNNs · Long-delay extrapolation across datasets, architectures, unit types, and training methods | retrieval accuracy | Helped |
| Add hidden-state speed regularization targeting slow points (ablation) | DeCu GRU on MNIST · Extremely long delays | retrieval accuracy | Helped |
| Add hidden-state speed regularization targeting class centers of mass (ablation) | Delayed classification RNNs · Extended delays across datasets, curricula, and unit types | retrieval accuracy | Helped |
| Use class centers of mass as regularization targets instead of detected slow points (ablation) | Delayed classification RNNs · Extended delays | retrieval accuracy | No clear effect |
| Add a new class that recruits an existing class's slow point | Vocabulary curriculum training · Accuracy of the existing class associated with the recruited slow point | class accuracy | Hurt |
| Train with vocabulary curriculum rather than delay curriculum | Delayed classification RNNs · Nominal-delay task | retrieval accuracy | No clear effect |
| Train without curricula (ablation) | Delayed template-matching RNNs · MNIST nominal-delay template matching | task performance | No clear effect |
| Increase slow-point speed | Delayed template-matching RNNs · Long-delay MNIST template matching | accuracy | Hurt |
Closest papers by meaning
- Dynamical phases of short-term memory mechanisms in RNNs (ICML 2025 · similarity 0.91)
- On the Curse of Memory in Recurrent Neural Networks: Approximation and Optimization Analysis (ICLR 2021 · similarity 0.90)
- Recurrent neural networks: vanishing and exploding gradients are not the end of the story (NeurIPS 2024 · similarity 0.90)
- Learning dynamics in linear recurrent neural networks (ICML 2025 · similarity 0.89)
- Revisiting Bi-Linear State Transitions in Recurrent Neural Networks (NeurIPS 2025 · similarity 0.89)
- Why Diffusion Models Don’t Memorize: The Role of Implicit Dynamical Regularization in Training (NeurIPS 2025 · similarity 0.89)
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.