Research map / Optimization / Optimizers
MomentumRNN: Integrating Momentum into Recurrent Neural Networks
Tan Nguyen, Richard Baraniuk, Andrea Bertozzi, Stanley Osher, Bao Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 a momentum state to LSTM | LSTM · PMNIST | test accuracy (%): 92.29 -> 94.72 | Improved |
| Add a momentum state to LSTM | LSTM · TIMIT test | MSE: 9.37 -> 5.87 | Improved |
| Add a momentum state to LSTM | LSTM · Copying, sequence length 2K | training loss: 0.01 -> 0.009 | Improved |
| Add a momentum state to LSTM | LSTM · Adding, sequence length 750 | training loss: 0.162 -> 0.162 | No clear effect |
| Add a momentum state to LSTM | LSTM · PTB word-level test | perplexity: 59.71 \pm 0.99 -> 58.62 \pm 0.22 | Improved |
| Add a momentum state to DTRIV | DTRIV · PMNIST | test accuracy (%): 96.62 \pm 0.12 -> 96.89 \pm 0.08 | Improved |
| Add a momentum state to DTRIV | DTRIV · TIMIT test | MSE: 1.87 \pm 0.17 -> 1.17 \pm 0.05 | Improved |
| Replace standard momentum updates with Adam-style recurrent updates | MomentumLSTM · PMNIST | test accuracy (%): 94.72 \pm 0.16 -> 95.38 \pm 0.19 | Improved |
| Use AdamLSTM or RMSPropLSTM instead of MomentumLSTM | LSTM · TIMIT test | MSE: 5.87 \pm 0.15 -> 8.69 \pm 0.14 (AdamLSTM); 9.17 \pm 0.33 (RMSPropLSTM) | Hurt |
| Use scheduled-restart momentum instead of constant momentum | MomentumLSTM · TIMIT test | MSE: 5.87 \pm 0.15 -> 5.83 \pm 0.10 | Improved |
| Use Adam-style recurrent updates instead of standard momentum | MomentumLSTM · Copying and adding tasks | training loss | Improved |
| Use AdamLSTM or RMSPropLSTM instead of LSTM | LSTM · PTB word-level test | perplexity: 59.71 \pm 0.99 -> 61.11 \pm 0.31 (AdamLSTM); 64.53 \pm 0.20 (RMSPropLSTM) | Hurt |
Closest papers by meaning
- Momentum Residual Neural Networks (ICML 2021 · similarity 0.89)
- MDN: Parallelizing Stepwise Momentum for Delta Linear Attention (ICML 2026 · similarity 0.88)
- Recurrent neural networks: vanishing and exploding gradients are not the end of the story (NeurIPS 2024 · similarity 0.87)
- The Marginal Value of Momentum for Small Learning Rate SGD (ICLR 2024 · similarity 0.87)
- Enhancing Optimizer Stability: Momentum Adaptation of The NGN Step-size (NeurIPS 2025 · similarity 0.86)
- In Search of Adam’s Secret Sauce (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.
