Research map / Architectures and efficiency / RNNs and normalization
State-Regularized Recurrent Neural Networks to Extract Automata and Explain Predictions
Cheng Wang, Carolin Lawrence, Mathias Niepert
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 state regularization with projection (sr-LSTM-p) | LSTM · Balanced parentheses (small train set), d=20, l<=1000 | error rate: 0.519 -> 0.380 | Improved |
| Add state regularization with projection (sr-LSTM-p) | LSTM · Balanced parentheses (large train set), d in [10,20], l<=200 | error rate: 0.341 -> 0.003 | Improved |
| Add state regularization without projection (sr-LSTM) | LSTM · Balanced parentheses (large train set), d=5, l<=200 | error rate: 0.002 -> 0.044 | Hurt |
| Add state regularization with projection (sr-LSTM-p) | LSTM · Palindrome test, max length 100 | error rate (%): 31.2 -> 10.5 | Improved |
| Add state regularization with projection (sr-LSTM-p) | LSTM · Palindrome test, max length 500 | error rate (%): 47.7 -> 29.8 | Improved |
| Add projection to state-regularized LSTM (ablation) | sr-LSTM · Palindrome test, max length 200 | error rate (%): 36.0 -> 16.7 | Improved |
| Add state regularization to LSTM | LSTM · IMDB | test error (%): 10.1 -> 9.2 | Improved |
| Use alternative sr-LSTM-p formulations instead of best one (ablation) | sr-LSTM-p (k=10) · IMDB | test error (%): 9.2 -> 11.1 | Hurt |
| Add state regularization to LSTM | LSTM · Sequential MNIST | test error (%): 2.3 -> 0.8 | Improved |
| Use alternative sr-LSTM-p formulation instead of best one (ablation) | sr-LSTM-p (k=100) · Sequential MNIST | test error (%): 0.8 -> 6.7 | Hurt |
| Reduce centroids from k=5 to k=2 (ablation) | sr-LSTM-p · Balanced parentheses (small train set), d=10, l<=200 | error rate: 0.138 -> 0.038 | Improved |
| Increase centroids from k=5 to k=50 (ablation) | sr-LSTM-p · Balanced parentheses (small train set), d=5, l<=200 | error rate: 0.015 -> 0.047 | Hurt |
Closest papers by meaning
- State-Regularized Recurrent Neural Networks (ICML 2019 · similarity 0.98)
- Decision-Guided Weighted Automata Extraction from Recurrent Neural Networks (AAAI 2021 · similarity 0.91)
- Extracting Automata from Recurrent Neural Networks Using Queries and Counterexamples (ICML 2018 · similarity 0.91)
- Representing Formal Languages: A Comparison Between Finite Automata and Recurrent Neural Networks (ICLR 2019 · similarity 0.90)
- Recurrent neural networks: vanishing and exploding gradients are not the end of the story (NeurIPS 2024 · similarity 0.88)
- M$^2$RNN: Non-Linear RNNs with Matrix-Valued States for Scalable Language Modeling (NeurIPS 2026 · similarity 0.88)
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.
