Research map / Architectures and efficiency / RNNs and normalization
Long Expressive Memory for Sequence Modeling
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 LEM instead of coRNN | sMNIST · sMNIST | test accuracy: 99.3% -> 99.5% | Helped |
| Use LEM instead of coRNN | psMNIST · psMNIST | test accuracy: 96.6% -> 96.6% | No clear effect |
| Use LEM instead of coRNN | nCIFAR-10 · nCIFAR-10 | test accuracy: 59.0% -> 60.5% | Helped |
| Use LEM instead of UnICORNN (2 layers) | EigenWorms · EigenWorms | test accuracy: 90.3% ± 3.0% -> 92.3% ± 1.8% | Helped |
| Use LEM instead of UnICORNN (3 layers) | Heart-rate prediction · Heart-rate prediction (PPG) | test L^2 error: 1.31 -> 0.85 | Helped |
| Use LEM instead of coRNN | FitzHugh-Nagumo prediction · FitzHugh-Nagumo | test L^2 error: 0.4 -> 0.2 | Helped |
| Use LEM instead of GRU | Google12 · Google12 | test accuracy: 95.2% -> 95.7% | Helped |
| Use LEM instead of LSTM | PTB character-level · PTB character-level | test bpc: 1.36 -> 1.25 | Helped |
| Use LEM instead of TARNN | PTB word-level · PTB word-level | test perplexity: 94.6 -> 72.8 | Helped |
| Add learnable multiscale time-step gates instead of fixing both gates to one tuned scale (ablation) | LEM on FitzHugh-Nagumo · FitzHugh-Nagumo | test L^2 error: 1.1 -> 0.2 | Helped |
| Add chrono initialization while keeping Δt fixed to 1 (ablation) | LEM on EigenWorms · EigenWorms | test accuracy: 57.9% ± 7.7% -> 88.2% ± 6.9% | Helped |
| Tune Δt without chrono initialization (ablation) | LEM on EigenWorms · EigenWorms | test accuracy: 57.9% ± 7.7% -> 92.3% ± 1.8% | Helped |
Closest papers by meaning
- AntisymmetricRNN: A Dynamical System View on Recurrent Neural Networks (ICLR 2019 · similarity 0.88)
- Stable Recurrent Models (ICLR 2019 · similarity 0.88)
- Kernel-Based Approaches for Sequence Modeling: Connections to Neural Methods (NeurIPS 2019 · similarity 0.88)
- State-Frequency Memory Recurrent Neural Networks (ICML 2017 · similarity 0.88)
- Recurrent neural networks: vanishing and exploding gradients are not the end of the story (NeurIPS 2024 · similarity 0.87)
- xLSTM: Extended Long Short-Term Memory (NeurIPS 2024 · 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.