Research map / Graph learning / Knowledge and temporal graphs
Knowledge-Aware Neuron Interpretation for Scene Classification
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 a knowledge-aware neuron interpretation framework using ConceptNet core concepts, concept filtering via KG embeddings, and model manipulation to explain and improve scene classifiers on ADE20k and Opensurfaces.
Problem
Explanations of scene classification predictions based on neuron concepts lack concept completeness, fail to merge semantically-equivalent concepts, and provide no way to verify or manipulate the original model's behavior.
Approach
Defines scene core concepts (scoping and identifier) from ConceptNet and dataset statistics, uses a MinMax-based NetDissect variant for per-image neuron concept assignment, merges concepts via KG embedding clustering (Concept Filtering), and uses core concepts for neuron ablation, concept-loss re-training, and SVM re-training on explanation metrics.
Key results
- Concept filtering with TransE and 168 clusters yields up to +26.3% IoU gain over non-clustered concepts on ResNet-18 (ADE20k).
- SVM re-training on ICC explanation metrics raises ADE20k accuracy from 52.96% to 67.11% over ResNet18 (about 26.7% relative improvement).
- SCC achieves the best false-prediction explanation scores, e.g. CM^FP 78.51% vs 43.04% for Top_10 under whole-layer selection.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use core concepts (SCC) instead of Top_10 as the concept reference (ablation) | Threshold-based neuron concepts · ADE20k false prediction explanation | CM^FP (%): 42.52 -> 78.03 | Helped |
| Use core concepts (SCC) instead of Top_10 as the concept reference (ablation) | Threshold-based neuron concepts · ADE20k false prediction explanation | DM^FP (%): 18,69 -> 87.38 | Helped |
| Use core concepts (SCC) instead of Top_10 as the concept reference (ablation) | Threshold-based neuron concepts · ADE20k false prediction explanation | SM^FP (%): 42.42 -> 72.13 | Helped |
| Replace unclustered concepts with Concept Filtering clusters (ablation) | Neuron interpretation · ADE20k | IoU gain (%): +26.31 | Helped |
| Use TransE embeddings rather than TransD, Dismult, or ProjE (ablation) | Concept Filtering · ADE20k neuron interpretation, 168 clusters | IoU gain (%): +26.31 | Mixed |
| Vary the number of concept clusters (ablation) | Concept Filtering · ADE20k neuron interpretation | IoU gain (%): +17.01; +18.11; +18.44; +21.78; +20.72; +20.90; +21.27; +25.29; +24.10 | Mixed |
| Use ICC rather than SCC (ablation) | SVM retraining via prediction-explanation features · ADE20k | accuracy (%): 66.54 -> 67.11 | Helped |
| Use ICC rather than SCC (ablation) | SVM retraining via prediction-explanation features · Opensurfaces | accuracy (%): 31.62 -> 32.27 | Helped |
| Add prediction-explanation metrics and hidden-state features via SVM (ablation) | ResNet18 · ADE20k | accuracy (%): 52.96 -> 67.11 | Helped |
| Add prediction-explanation metrics and hidden-state features via SVM (ablation) | ResNet18 · Opensurfaces | accuracy (%): 29.26 -> 32.27 | Helped |
| Disable negative neurons identified using core concepts (ablation) | Scene classification model · ADE20k and Opensurfaces | accuracy | Helped |
| Disable positive neurons identified using core concepts (ablation) | Scene classification model · ADE20k and Opensurfaces | accuracy | Hurt |
Surprising findings
- ICC performs better for re-training and neuron identification, while SCC is better for explaining false predictions.
Limitations
- Concept alignment between ADE20k and ConceptNet requires fuzzy matching for a fraction of concepts.
- Opensurfaces concepts such as material properties yield limited scene differentiation, reducing the manipulation gains.
- Experiments on false-prediction explanation across models use only 1000 random ADE20k samples due to time cost.
Experimental setup
Datasets: ADE20k Opensurfaces ConceptNet
Benchmarks: ADE20k Opensurfaces
Metrics: CM SM DM IoU gain accuracy
Baselines: ConceptSHAP CLIP-Dissect Top_10 concepts ResNet18
Builds on: NetDissect CLIP-Dissect ConceptSHAP ConceptNet TransE
Closest papers by meaning
- NeuronEye: Query-Guided Visual Concept Activation for Vision-Language Reasoning (NeurIPS 2026 · similarity 0.88)
- LINE: LLM-based Iterative Neuron Explanations for Vision Models (NeurIPS 2026 · similarity 0.88)
- Automated Natural Language Explanation of Deep Visual Neurons with Large Models (Student Abstract) (AAAI 2024 · similarity 0.86)
- Show and Tell: Visually Explainable Deep Neural Nets via Spatially-Aware Concept Bottleneck Models (CVPR 2025 · similarity 0.86)
- Concept-RuleNet: Grounded Multi-Agent Neurosymbolic Reasoning in Vision Language Models (AAAI 2026 · similarity 0.86)
- Understanding Distributed Representations of Concepts in Deep Neural Networks without Supervision (AAAI 2024 · 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.