Research map / Architectures and efficiency / RNNs and normalization
Fast-Slow 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 the FS-LSTM-2 architecture instead of the strongest listed previous model | Character-level language modeling · Penn Treebank | BPC: 1.214 -> 1.190 | Helped |
| Use four Fast cells instead of two (ablation) | FS-LSTM · Penn Treebank | BPC: 1.190 -> 1.193 | Hurt |
| Use the FS-LSTM-2 architecture instead of the strongest listed previous neural model | Character-level language modeling · enwik8 | BPC: 1.27 -> 1.290 | Helped |
| Use four Fast cells instead of two (ablation) | FS-LSTM · enwik8 | BPC: 1.290 -> 1.277 | Helped |
| Increase model size (ablation) | FS-LSTM-4 · enwik8 | BPC: 1.277 -> 1.245 | Helped |
| Ensemble two Large FS-LSTM-4 models (ablation) | Large FS-LSTM-4 · enwik8 | BPC: 1.245 -> 1.198 | Helped |
| Use an ensemble of two FS-LSTM models rather than cmix | enwik8 compression comparison · enwik8 | BPC: 1.23 -> 1.20 | Helped |
| Use FS-LSTM rather than sequential-LSTM | Architecture comparison · enwik8 (20-epoch architecture comparison) | BPC: 1.58 -> 1.49 | Helped |
| Use FS-LSTM rather than stacked-LSTM | Architecture comparison · enwik8 (20-epoch architecture comparison) | BPC: 1.61 -> 1.49 | Helped |
| Add further slower hierarchical layers (ablation) | FS-RNN · Tasks considered in the paper | test performance | Hurt |
| Increase the model size (ablation) | FS-LSTM · Penn Treebank | BPC | No clear effect |
Closest papers by meaning
- LightRNN: Memory and Computation-Efficient Recurrent Neural Networks (NeurIPS 2016 · similarity 0.90)
- Variable Computation in Recurrent Neural Networks (ICLR 2017 · similarity 0.90)
- FlashRNN: Unlocking Parallel Training of Nonlinear RNNs for Large Language Models (ICLR 2026 · similarity 0.89)
- State-Regularized Recurrent Neural Networks (ICML 2019 · similarity 0.89)
- Resurrecting Recurrent Neural Networks for Long Sequences (ICML 2023 · similarity 0.89)
- M$^2$RNN: Non-Linear RNNs with Matrix-Valued States for Scalable Language Modeling (NeurIPS 2026 · 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.