Research map / Architectures and efficiency / RNNs and normalization
Do RNN and LSTM have Long Memory?
Jingyu Zhao, Feiqing Huang, Jia Lv, Yanjie Duan, Zhen Qin, Guodong Li, Guangjian Tian
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 the memory filter to RNN | RNN · ARFIMA, DJI, Traffic, and Tree | RMSE: MRNN: 1.1620 -> 1.0880; 0.2605 -> 0.2487; 336.44 -> 333.72; 0.2871 -> 0.2818 | Improved |
| Add the memory filter to LSTM cell states | LSTM · ARFIMA, DJI, Traffic, and Tree | RMSE: 1.1340 -> 1.1490 | Hurt |
| Use a fixed memory parameter instead of dynamic d (ablation) | MRNN · ARFIMA, DJI, Traffic, and Tree | RMSE: 1.0880 -> 1.1010; 0.2487 -> 0.2472; 333.72 -> 333.36; 0.2818 -> 0.2822 | Mixed |
| Use a fixed memory parameter instead of dynamic d (ablation) | MLSTM · ARFIMA, DJI, Traffic, and Tree | RMSE: 1.1490 -> 1.1580; 0.2531 -> 0.2540; 337.83 -> 337.78; 0.2859 -> 0.2859 | Mixed |
| Increase truncation length K (ablation) | MRNN and MRNNF · ARFIMA, DJI, Traffic, and Tree | RMSE | Improved |
| Increase truncation length K (ablation) | MLSTM and MLSTMF · ARFIMA, DJI, Traffic, and Tree | RMSE | Hurt |
| Use K=25 rather than the default K=100 (ablation) | MLSTM and MLSTMF · ARFIMA and Traffic | RMSE | Improved |
| Add a fixed memory filter to RNN | RNN · CMU-MOSI | MAE: 1.5028 -> 1.4953 | Improved |
| Add a fixed memory filter to LSTM cell states | LSTM · CMU-MOSI | MAE: 1.4978 -> 1.4972 | No clear effect |
| Use MLSTMF50 as the first layer, with LSTM as the second | Two-layer sentiment classifier · Paper Reviews | Accuracy: 0.3021 -> 0.3110 | Improved |
| Use MLSTMF50 as the first layer, with LSTM as the second | Two-layer sentiment classifier · Paper Reviews | Cross-entropy loss: 1.5752 -> 1.4758 | Improved |
| Use memory-augmented models instead of RNN on a short-memory synthetic series | MRNN(F) and MLSTM(F) · RNN synthetic dataset | Forecast error | No clear effect |
Closest papers by meaning
- Training biologically plausible recurrent neural networks on cognitive tasks with long-term dependencies (NeurIPS 2023 · similarity 0.88)
- Implicit Bias of Linear RNNs (ICML 2021 · similarity 0.88)
- Recurrent Orthogonal Networks and Long-Memory Tasks (ICML 2016 · similarity 0.87)
- Understanding and Controlling Memory in Recurrent Neural Networks (ICML 2019 · similarity 0.87)
- Recurrent neural networks: vanishing and exploding gradients are not the end of the story (NeurIPS 2024 · similarity 0.87)
- Capacity and Trainability in Recurrent Neural Networks (ICLR 2017 · similarity 0.87)
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.
