Research map / Architectures and efficiency / RNNs and normalization
Improving the Gating Mechanism of 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 |
|---|---|---|---|
| Add the refine gate to uniform gate initialization (ablation) | LSTM forget gate · WikiText-103 test | perplexity: 34.9 -> 34.6 | Helped |
| Use uniform gate initialization instead of standard initialization (ablation) | LSTM forget gate · WikiText-103 test | perplexity: 35.8 -> 34.9 | Helped |
| Add the refine gate to standard initialization (ablation) | LSTM forget gate · pMNIST validation | accuracy: 94.77 -> 95.84 | Helped |
| Add uniform gate initialization and a refine gate (ablation) | LSTM forget gate · sCIFAR-10 validation | accuracy: 63.24 -> 71.05 | Helped |
| Use uniform gate initialization without a refine gate (ablation) | LSTM forget gate · Copy task, sequence length 500 | loss | Mixed |
| Combine a refine gate with a broad range of gate values (ablation) | LSTM forget gate · Copy task, sequence length 500 | loss | Helped |
| Use a refine gate with standard initialization instead of combining it with a broad initialization range (ablation) | LSTM forget gate · Copy task, sequence length 500 | loss | Hurt |
| Use uniform initialization and refine gates instead of standard gates (ablation) | LSTM · Adding task, sequence length 2000 | mean squared error | Helped |
| Replace chrono initialization with uniform gate initialization (ablation) | LSTM forget gate · WikiText-103 | perplexity: 36.4 -> 34.9 | Helped |
| Manually set chrono initialization T_max to 8 (ablation) | LSTM forget gate · WikiText-103 test | perplexity: 36.4 -> 36.1 | No clear effect |
| Add Zoneout to URLSTM (ablation) | URLSTM · sCIFAR-10 | test accuracy: 71.00 -> 74.34 | Helped |
| Add UR gates to GRU and Zoneout (ablation) | GRU · sCIFAR-10 | test accuracy: +3.1 | Helped |
Closest papers by meaning
- On the Importance of Gating: Memorization vs. In-Context Learning in State Space Models (NeurIPS 2026 · similarity 0.89)
- Gated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-Free (NeurIPS 2025 · similarity 0.85)
- Rethinking Neural Nonlinearity as Gating (NeurIPS 2026 · similarity 0.85)
- Dynamical phases of short-term memory mechanisms in RNNs (ICML 2025 · similarity 0.85)
- Recurrent neural networks: vanishing and exploding gradients are not the end of the story (NeurIPS 2024 · similarity 0.85)
- Learning Delays in Spiking Neural Networks using Dilated Convolutions with Learnable Spacings (ICLR 2024 · similarity 0.84)
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.