Research map / Architectures and efficiency / Pruning and efficient tuning
Discovering Salient Neurons in deep NLP models
Nadir Durrani, Fahim Dalvi, 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 |
|---|---|---|---|
| keep LCA top 20% neurons instead of random 20% (ablation) | BERT · POS | accuracy: 28.45 -> 90.16 | Improved |
| keep LCA top 20% neurons instead of bottom 20% (ablation) | XLNet · SEM | accuracy: 25.37 -> 90.70 | Improved |
| rank neurons by LCA instead of weight magnitude (ablation) | RoBERTa · POS | accuracy: 71.06 -> 91.93 | Improved |
| rank neurons by LCA instead of random selection (ablation) | BERT · Chunking | accuracy: 75.83 -> 89.01 | Improved |
| retrain classifier on top 5% neurons instead of all neurons | XLNet · POS | accuracy: 96.19 -> 96.49 | No clear effect |
| rank neurons with LCA instead of Probeless (ablation) | BERT · Chunking (10% neurons) | accuracy: 90.36 -> 94.99 | Improved |
| rank neurons with LCA instead of Probeless (ablation) | BERT · POS (5% neurons) | accuracy: 95.45 -> 95.92 | No clear effect |
| select top neurons only instead of all neurons | BERT · POS | selectivity: 14.49 -> 31.84 | Improved |
| retrain on random neurons instead of top neurons (ablation) | BERT · POS (5% neurons) | accuracy: 95.92 -> 94.12 | Hurt |
| retrain on bottom neurons instead of top neurons after QNLI fine-tuning | XLNet · POS | accuracy: 96.49 -> 10.39 | Hurt |
| fine-tune on GLUE task (MRPC) before probing bottom neurons | BERT · POS | accuracy: 93.20 -> 93.13 | No clear effect |
| use top neurons in multilingual XLM-R instead of all neurons | XLM-R · POS English | accuracy: 94.70 -> 94.17 | No clear effect |
Closest papers by meaning
- Journey to the Center of the Knowledge Neurons: Discoveries of Language-Independent Knowledge Neurons and Degenerate Knowledge Neurons (AAAI 2024 · similarity 0.88)
- Far from the Shallow: Brain-Predictive Reasoning Embedding through Residual Disentanglement (NeurIPS 2025 · similarity 0.88)
- Towards Neuron Attributions in Multi-Modal Large Language Models (NeurIPS 2024 · similarity 0.88)
- The Emergence of Abstract Thought in Large Language Models Beyond Any Language (NeurIPS 2025 · similarity 0.88)
- What does the Knowledge Neuron Thesis Have to do with Knowledge? (ICLR 2024 · similarity 0.88)
- How do Large Language Models Handle Multilingualism? (NeurIPS 2024 · similarity 0.87)
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.
