Research map / Architectures and efficiency / RNNs and normalization
Weighted Automata Extraction from Recurrent Neural Networks via Regression on State Spaces
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 regression-based equivalence queries with M=5 instead of breadth-first search with n=5000 | WFA extraction from 2-layer LSTM RNNs · More WFA-like RNNs | MSE (10^-4): 4.21 -> 4.40 | Hurt |
| Use regression-based equivalence queries with M=2 instead of breadth-first search with n=5000 | WFA extraction from 2-layer LSTM RNNs · Realistic RNNs with nonuniform input domains | MSE (10^-4): 6.38 -> 6.02 | Helped |
| Increase the concentration threshold from M=2 to M=5 (ablation) | Regression-based WFA extraction · More WFA-like RNNs | MSE (10^-4): 5.45 -> 4.40 | Helped |
| Increase the concentration threshold from M=2 to M=5 (ablation) | Regression-based WFA extraction · Realistic RNNs with nonuniform input domains | MSE (10^-4): 6.02 -> 6.90 | Hurt |
| Use a larger concentration threshold, M=5 rather than M=2 (ablation) | Regression-based WFA extraction · wparen weighted balanced-parentheses task | recognizable balanced-parentheses depth: depth one -> depth two | Helped |
| Use Gaussian process regression rather than kernel ridge regression for the configuration abstraction function (ablation) | Regression-based WFA extraction · Experiments across the reported RNN tasks | extraction performance | Helped |
| Replace RNN inference with extracted WFA inference | Target RNNs and their extracted WFAs · Inference-time experiments | inference speed: about 1,300 times faster | Helped |
Closest papers by meaning
- Decision-Guided Weighted Automata Extraction from Recurrent Neural Networks (AAAI 2021 · similarity 0.92)
- Extracting Automata from Recurrent Neural Networks Using Queries and Counterexamples (ICML 2018 · similarity 0.92)
- Learning Deterministic Weighted Automata with Queries and Counterexamples (NeurIPS 2019 · similarity 0.91)
- Learning Useful Representations of Recurrent Neural Network Weight Matrices (ICML 2024 · similarity 0.85)
- Recurrent neural networks: vanishing and exploding gradients are not the end of the story (NeurIPS 2024 · similarity 0.83)
- M$^2$RNN: Non-Linear RNNs with Matrix-Valued States for Scalable Language Modeling (NeurIPS 2026 · similarity 0.83)
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.