Research map / Architectures and efficiency / RNNs and normalization
Neuron Shapley: Discovering the Responsible Neurons
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 Shapley scores instead of neuron conductance to pick filters to remove | Inception-v3 · ImageNet accuracy collapse | filters needed to reach random accuracy: 30 vs twice as many for conductance | Helped |
| use Shapley instead of conductance to select unfair filters to remove | SqueezeNet · PPB | average accuracy (%): 88.7 -> 91.7 | Helped |
| zero out most negative Shapley filters | SqueezeNet · PPB black female | accuracy (%): 54.7 -> 81.9 | Helped |
| remove top 16 adversarial-Shapley filters | Inception-v3 · ImageNet PGD attack | adversary success rate (%): nearly 100 -> 0.1 | Helped |
| remove top 16 adversarial-Shapley filters | Inception-v3 · ImageNet clean images | clean accuracy (%): 74 -> 67 | Hurt |
| use Shapley instead of conductance for vulnerable filters | Inception-v3 · ImageNet clean accuracy after removal | clean accuracy (%): 53.2 -> 67 | Helped |
| use Shapley instead of conductance for class-specific filters | Inception-v3 · Carousel class top-10 | class accuracy after removal (%): 64 -> 20 | Helped |
| add multi-armed bandit adaptive sampling to MC-Shapley (ablation) | MC-Shapley · SqueezeNet | samples required: 8.7% of MC-Shapley samples | Helped |
| add early truncation (ablation) | Shapley estimation · Inception-v3 | compute | Helped |
| train with filter dropout (p=0.5) (ablation) | SqueezeNet · CelebA | robustness to removing top-30 filters: test acc 95.0% vs 98.0% | Helped |
Closest papers by meaning
- Know Thyself by Knowing Others: Learning Neuron Identity from Population Context (NeurIPS 2025 · similarity 0.79)
- NeurFlow: Interpreting Neural Networks through Neuron Groups and Functional Interactions (ICLR 2025 · similarity 0.78)
- Neurons as Detectors of Coherent Sets in Sensory Dynamics (NeurIPS 2025 · similarity 0.78)
- Evaluating Neuron Explanations: A Unified Framework with Sanity Checks (ICML 2025 · similarity 0.78)
- Neuron Activation Coverage: Rethinking Out-of-distribution Detection and Generalization (ICLR 2024 · similarity 0.78)
- The Expressive Leaky Memory Neuron: an Efficient and Expressive Phenomenological Neuron Model Can Solve Long-Horizon Tasks. (ICLR 2024 · similarity 0.78)
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.