Research map / Architectures and efficiency / RNNs and normalization
Neural Speed Reading via Skim-RNN
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 |
|---|---|---|---|
| Replace standard LSTM with Skim-LSTM | LSTM text classifiers · SST, Rotten Tomatoes, IMDb, and AGNews | accuracy: 86.4 -> 86.4; 82.5 -> 84.2; 91.1 -> 91.2; 93.5 -> 93.6 | Mixed |
| Replace standard LSTM with Skim-LSTM | LSTM text classifiers · SST, Rotten Tomatoes, IMDb, and AGNews | Flop reduction: 2.4x; 2.1x; 4.7x; 1.4x | Helped |
| Replace standard LSTM with Skim-LSTM | CBT question answering model · CBT-NE and CBT-CN | accuracy: 49.0 -> 51.4; 54.9 -> 56.3 | Helped |
| Replace LSTM with Skim-LSTM | LSTM+Attention · SQuAD | F1: 47.8 -> 48.6 | Helped |
| Replace LSTM with Skim-LSTM | BiDAF · SQuAD | F1: 77.3% -> 77.1% | Hurt |
| Increase γ from 0.01 to 0.02 (ablation) | Skim-LSTM · IMDb | Flop reduction: 4.7x -> 9.5x | Helped |
| Increase γ from 0.01 to 0.02 (ablation) | Skim-LSTM · IMDb | accuracy: 91.2 -> 90.9 | Hurt |
| Increase γ from 0.01 to 0.02 (ablation) | Skim-LSTM · CBT-NE | accuracy: 50.9 -> 51.4 | Helped |
| Increase γ from 0.01 to 0.02 (ablation) | Skim-LSTM · CBT-CN | accuracy: 56.3 -> 51.4 | Hurt |
| Train Skim-LSTM from scratch rather than using partial standard-LSTM pretraining (ablation) | Skim-LSTM+Attention · SQuAD | F1 | Hurt |
| Use full rather than half standard-LSTM pretraining (ablation) | Skim-LSTM+Attention · SQuAD | F1: 75.5; 70.7 | Hurt |
| Replace skimming with skipping (set d'=0) (ablation) | Skim-LSTM+Attention · SQuAD | F1 | Hurt |
Closest papers by meaning
- Resurrecting Recurrent Neural Networks for Long Sequences (ICML 2023 · similarity 0.88)
- Neural Speed Reading with Structural-Jump-LSTM (ICLR 2019 · similarity 0.88)
- Time Adaptive Recurrent Neural Network (CVPR 2021 · similarity 0.87)
- Recurrent neural networks: vanishing and exploding gradients are not the end of the story (NeurIPS 2024 · similarity 0.86)
- FlashRNN: I/O-Aware Optimization of Traditional RNNs on modern hardware (ICLR 2025 · similarity 0.85)
- M$^2$RNN: Non-Linear RNNs with Matrix-Valued States for Scalable Language Modeling (NeurIPS 2026 · similarity 0.85)
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.