Research map / Architectures and efficiency / RNNs and normalization
Reverse engineering recurrent networks for sentiment classification reveals line attractor dynamics
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 recurrent architectures instead of bag-of-words classification | Sentiment classification models · Yelp 2015 | accuracy: 93.37% (Bag of words); 92.96% (Vanilla RNN); 95.67% (Update Gate RNN); 95.84% (GRU); 95.05% (LSTM) | Mixed |
| Use recurrent architectures instead of bag-of-words classification | Sentiment classification models · IMDB | accuracy: 88.53% (Bag of words); 87.08% (Vanilla RNN); 87.96% (Update Gate RNN); 86.86% (GRU); 86.93% (LSTM) | Hurt |
| Use recurrent architectures instead of bag-of-words classification | Sentiment classification models · Stanford Sentiment | accuracy: 79.74% (Bag of words); 78.09% (Vanilla RNN); 77.74% (Update Gate RNN); 80.25% (GRU); 80.09% (LSTM) | Mixed |
| Linearize the trained LSTM dynamics around fixed points instead of using the full nonlinear model | LSTM sentiment classifier · Yelp reviews | accuracy | Hurt |
| Train a linear model directly rather than linearizing a trained nonlinear model | Sentiment classification · Yelp reviews | accuracy: around 3% worse than the LSTM | Helped |
| Train recurrent networks on sentiment classification | RNN hidden-state representations · Yelp reviews | variance captured by top dimensions: ~90% | Helped |
Closest papers by meaning
- The geometry of integration in text classification RNNs (ICLR 2021 · similarity 0.91)
- How recurrent networks implement contextual processing in sentiment analysis (ICML 2020 · similarity 0.90)
- Reverse engineering recurrent neural networks with Jacobian switching linear dynamical systems (NeurIPS 2021 · similarity 0.88)
- Learning dynamics in linear recurrent neural networks (ICML 2025 · similarity 0.87)
- Learning Useful Representations of Recurrent Neural Network Weight Matrices (ICML 2024 · similarity 0.86)
- Mechanistic Interpretability of RNNs emulating Hidden Markov Models (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.