Research map / Architectures and efficiency / RNNs and normalization
SPICE: Simple Polysemantic feature Interpretation via Clustering-based Explanations
Sehyun Lee, Dahee Kwon, Damin Lee, Jaesik Choi
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 adaptive per-neuron cluster discovery instead of fixed K=2 (Input x Gradient) (ablation) | SPICE · ViT-B/16 B11 | separability: 1.027 -> 1.308 | Improved |
| add per-dimension attribution normalization on top of adaptive K (ablation) | SPICE · ViT-B/16 B11 | separability: 1.308 -> 1.577 | Improved |
| replace CRP with Input x Gradient under fixed K=2 (ablation) | PURE-style clustering · ViT-B/16 B11 | separability: 1.036 -> 1.027 | No clear effect |
| use constant K in {2,3,4} instead of adaptive K (ablation) | SPICE · ViT-B/16 B11 | separability: 1.577 -> 1.027/1.025/1.026 | Hurt |
| add feature-wise normalization (ablation) | SPICE · ResNet-50 L4.2 | separability: 1.017 -> 1.293 | Improved |
| add feature-wise normalization (ablation) | SPICE · ViT-B B2 | separability: 1.065 -> 1.137 | Improved |
| add feature-wise normalization (ablation) | SPICE · ViT-B B11 | separability: 1.507 -> 1.577 | Improved |
| use LeGrad instead of Input x Gradient attribution (ablation) | SPICE · ViT-B/16 B11 | separability: 1.577 -> 1.493 | No clear effect |
| use static global tau>=0.9 instead of adaptive tau (ablation) | SPICE · ViT-B B6 | separability / coverage: score 1.25 -> 2.18, coverage 100/100 -> 7/100 | Mixed |
| use SPICE instead of LE in deep layers | LE baseline · ViT-B B11 | separability: 1.357 -> 1.577 | Improved |
| use SPICE instead of LE in early layers | LE baseline · ViT-B B2 | separability: 1.199 -> 1.137 | Hurt |
| use SPICE instead of LE in DenseNet L4.16 | LE baseline · DenseNet L4.16 | separability: 1.191 -> 1.033 | Hurt |
Closest papers by meaning
- Capturing Polysemanticity with PRISM: A Multi-Concept Feature Description Framework (NeurIPS 2025 · similarity 0.88)
- CoE: Chain-of-Explanation via Automatic Visual Concept Circuit Description and Polysemanticity Quantification (CVPR 2025 · similarity 0.88)
- Unveiling the Unseen: Identifiable Clusters in Trained Depthwise Convolutional Kernels (ICLR 2024 · similarity 0.86)
- Beyond Interpretability: The Gains of Feature Monosemanticity on Model Robustness (ICLR 2025 · similarity 0.86)
- LINE: LLM-based Iterative Neuron Explanations for Vision Models (NeurIPS 2026 · similarity 0.85)
- NeurFlow: Interpreting Neural Networks through Neuron Groups and Functional Interactions (ICLR 2025 · 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.
