Research map / Label-efficient and robust learning / Recognition and long-tail
WWW: A Unified Framework for Explaining What Where and Why of Neural Networks by Interpretation of Neuron Concepts
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes WWW, a training-free framework that names neuron concepts via adaptive cosine similarity with CLIP and combines neuron activation maps with Taylor-approximated Shapley values to give what/where/why explanations plus a heatmap-similarity uncertainty signal.
Problem
Neural network explanations usually cover only one or two of 'what' (neuron concept naming), 'where' (input region), and 'why' (reasoning behind a decision); neuron-naming methods are global and not per-sample.
Approach
Adaptive cosine similarity (CLIP image-concept similarity minus a base-template similarity) with adaptive thresholding names each neuron's concepts, and neuron activation maps weighted by Shapley-based neuron contributions produce localized per-sample concept heatmaps; similarity between class-level and sample-level heatmaps serves as an uncertainty estimate.
Key results
- On ResNet-50 final-layer neurons with WordNet concept set, WWW reaches F1 0.4197 versus 0.3647 for CLIP-Dissect and hit rate 0.645 versus 0.456.
- On ResNet-18 Places365 with Broden concepts, WWW reaches F1 0.2242 versus 0.1096 for CLIP-Dissect.
- Ablation on ResNet-50 shows adaptive cosine similarity raises F1 from 0.3265 (no template) to 0.4197.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use WWW instead of CLIP-Dissect | ResNet-50 · ImageNet-1k, Wordnet (80k) concepts | F1-score: 0.3647 ± 0.014 -> 0.4197 ± 0.012 | Helped |
| use WWW instead of CLIP-Dissect | ResNet-50 · ImageNet-1k, Wordnet (80k) concepts | Hit Rate: 0.456 -> 0.645 | Helped |
| use WWW instead of CLIP-Dissect | ResNet-50 · ImageNet-1k, Broden (1.2k) concepts | F1-score: 0.0328 ± 0.003 -> 0.0645 ± 0.007 | Helped |
| use WWW instead of CLIP-Dissect | ResNet-50 · ImageNet-1k, Broden (1.2k) concepts | Hit Rate: 0.108 -> 0.091 | Hurt |
| use WWW instead of CLIP-Dissect | ViT-B/16 · ImageNet-1k, Wordnet (80k) concepts | F1-score: 0.3679 ± 0.014 -> 0.4166 ± 0.012 | Helped |
| use WWW instead of CLIP-Dissect | ViT-B/16 · ImageNet-1k, Wordnet (80k) concepts | Hit Rate: 0.460 -> 0.634 | Helped |
| use WWW instead of CLIP-Dissect | ViT-B/16 · ImageNet-1k, Broden (1.2k) concepts | F1-score: 0.0319 ± 0.003 -> 0.0651 ± 0.007 | Helped |
| use WWW instead of CLIP-Dissect | ViT-B/16 · ImageNet-1k, Broden (1.2k) concepts | Hit Rate: 0.106 -> 0.091 | Hurt |
| use WWW instead of CLIP-Dissect | ResNet-18 · Places365, Wordnet (80k) concepts | F1-score: 0.1041 ± 0.016 -> 0.1106 ± 0.015 | Helped |
| use WWW instead of CLIP-Dissect | ResNet-18 · Places365, Broden (1.2k) concepts | F1-score: 0.1096 ± 0.008 -> 0.2242 ± 0.021 | Helped |
| add a base template (ablation) | WWW concept discovery on ResNet-50 · ImageNet-1k, Wordnet (80k) concepts | F1-score: 0.3265 ± 0.009 -> 0.3361 ± 0.009 | Helped |
| add a base template (ablation) | WWW concept discovery on ResNet-50 · ImageNet-1k, Wordnet (80k) concepts | mpnet cos: 0.6123 ± 0.007 -> 0.6075 ± 0.007 | Hurt |
Surprising findings
- On ImageNet with ImageNet (1k) concepts, WWW's CLIP cos (0.9325) is slightly below CLIP-Dissect (0.9340), so gains appear only for larger or different concept sets.
Limitations
- Uncertainty evaluation is shown only qualitatively via a rejection test on one setting, with limited comparison.
- On the ImageNet-1k concept set WWW does not beat CLIP-Dissect on CLIP cos, and the paper notes performance depends on the concept set.
- Concept sensitivity alpha must be tuned to maximize F1 on the target setting.
Experimental setup
Datasets: ImageNet-1k Places365 WordNet nouns Broden
Benchmarks: ImageNet-1k validation set Places365 test set
Metrics: CLIP cos mpnet cos F1-score Hit Rate
Baselines: Network Dissection MILAN FALCON CLIP-Dissect CRAFT
Builds on: CLIP-Dissect Network Dissection CRAFT Taylor approximation of Shapley value
Closest papers by meaning
- Evaluating Neuron Explanations: A Unified Framework with Sanity Checks (ICML 2025 · similarity 0.88)
- CoSy: Evaluating Textual Explanations of Neurons (NeurIPS 2024 · similarity 0.85)
- Neuron Explanations for Conformal Prediction (Student Abstract) (AAAI 2025 · similarity 0.84)
- NeurFlow: Interpreting Neural Networks through Neuron Groups and Functional Interactions (ICLR 2025 · similarity 0.84)
- Understanding Distributed Representations of Concepts in Deep Neural Networks without Supervision (AAAI 2024 · similarity 0.84)
- Neurons to Words: A Novel Method for Automated Neural Network Interpretability and Alignment (AAAI 2025 · 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.