Research map / Label-efficient and robust learning / Recognition and long-tail
Evaluating Neuron Interpretation Methods of NLP Models
Yimin Fan, Fahim Dalvi, Nadir Durrani, Hassan Sajjad
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 |
|---|---|---|---|
| use Probeless instead of Random neuron selection | neuron ranking · POS BERT | NeuronVote: 0.021 -> 0.532 | Improved |
| use Probeless instead of IoU (both corpus-based) | neuron ranking · POS BERT | NeuronVote: 0.365 -> 0.532 | Improved |
| use Probeless instead of LCA | neuron ranking · POS BERT layer 12 | NeuronVote: 0.391 -> 0.515 | Improved |
| use Probeless instead of Lasso | neuron ranking · POS XLM-R layer 12 | NeuronVote: 0.267 -> 0.497 | Improved |
| use Ridge instead of Lasso at last layer of RoBERTa | neuron ranking · POS RoBERTa layer 12 | NeuronVote: 0.350 -> 0.510 | Improved |
| move from layer 1 or 6 to last layer 12 | LCA and Lasso on BERT · POS | NeuronVote: 0.544 -> 0.391 | Hurt |
| use classifier accuracy as evaluation metric instead of compatibility | evaluation metric · POS BERT 30 neurons | accuracy: 0.92 -> 0.95 | Hurt |
| use random neuron selection with more neurons (ablation) | classifier accuracy metric · POS BERT | accuracy: 0.81 -> 0.92 | Improved |
| replace consensus of other methods with Lasso only when comparing Probeless vs LCA (ablation) | Probeless · POS BERT | NeuronVote: 0.449 vs 0.516 | Hurt |
| compare Probeless vs Lasso under all consensus subsets (ablation) | Probeless · POS BERT | NeuronVote: 0.382 vs 0.324 | Improved |
| compare Probeless vs IoU under all consensus subsets (ablation) | Probeless · POS BERT | NeuronVote: 0.426 vs 0.221 | Improved |
| use Gaussian instead of other methods | neuron ranking · POS BERT | NeuronVote: 0.169 vs 0.532 | Hurt |
Closest papers by meaning
- Evaluating Neuron Explanations: A Unified Framework with Sanity Checks (ICML 2025 · similarity 0.90)
- Automated Natural Language Explanation of Deep Visual Neurons with Large Models (Student Abstract) (AAAI 2024 · similarity 0.87)
- AND: Audio Network Dissection for Interpreting Deep Acoustic Models (ICML 2024 · similarity 0.86)
- Neurons to Words: A Novel Method for Automated Neural Network Interpretability and Alignment (AAAI 2025 · similarity 0.85)
- NeurFlow: Interpreting Neural Networks through Neuron Groups and Functional Interactions (ICLR 2025 · similarity 0.85)
- Do LLMs Signal When They’re Right? Evidence from Neuron Agreement (ICML 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.
