Research map / Probabilistic and scientific modeling / Molecules and proteins
Towards Human-Understandable Multi-Dimensional Concept Discovery
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 HU-MCD, which combines SAM-based segmentation for concept discovery, a CNN layer-masking scheme, and MCD's completeness-aware subspace decomposition to yield more understandable and faithful concept explanations for ResNet50 on ImageNet.
Problem
Concept-based XAI methods trade off human understandability of discovered concepts against faithful quantification of their importance; existing methods such as ACE and MCD either distort segments via inpainting/rescaling or lack understandability evaluation.
Approach
Use SAM segments as concept candidates, mask-aware CNN propagation (mask applied from the first layer onward, avoiding inpainting and rescaling) to embed them, then apply MCD's multi-dimensional subspace clustering and completeness-based local and global relevance scoring.
Key results
- In the human subject study (41 participants), HU-MCD reached 70.24% prediction accuracy versus 42.93% for ACE and 31.22% for MCD.
- HU-MCD had 67.12% recognizable concepts versus 45.66% (ACE) and 50.34% (MCD).
- In C-Deletion and C-Insertion, HU-MCD outperformed ACE and MCD, with C-Insertion near random for ACE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SAM-guided segmentation and CNN-specific input masking with the MCD framework | MCD (HU-MCD) · Human subject study | prediction accuracy: 42.93% -> 70.24%; 31.22% -> 70.24% | Helped |
| Use SAM-guided segmentation and CNN-specific input masking with the MCD framework | MCD (HU-MCD) · Human subject study | recognizable concepts: 45.66% -> 67.12%; 50.34% -> 67.12% | Helped |
| Use SAM-guided segmentation and CNN-specific input masking with the MCD framework | MCD (HU-MCD) · Human subject study | inner-concept description similarity: 0.39 -> 0.49; 0.41 -> 0.49 | Helped |
| Use SAM-guided segmentation and CNN-specific input masking with the MCD framework | MCD (HU-MCD) · Human subject study | intra-concept description similarity: 0.29 -> 0.28; 0.38 -> 0.28 | Helped |
| Use HU-MCD concept importance scores | ImageNet1k C-Deletion and C-Insertion evaluation · ImageNet1k, ten CIFAR10-like classes | average prediction accuracy during concept deletion and insertion | Helped |
Surprising findings
- MCD showed a large gap between prediction accuracy and recognizability because its concepts were highly similar within a class, making them hard to distinguish.
Limitations
- Non-coherent concepts can arise from segmentation errors, clustering inaccuracies, or similarity metric limits.
- Evaluation limited to visual data and ten ImageNet classes.
- Concept relevance is restricted to layers followed only by linear operations.
Experimental setup
Datasets: ImageNet1k
Benchmarks: ImageNet1k (ten CIFAR-10-like classes)
Metrics: prediction accuracy in task-prediction human study percentage of recognizable concepts inner- and intra-concept description similarity (GloVe) C-Deletion C-Insertion
Baselines: ACE MCD Random
Builds on: Multi-Dimensional Concept Discovery ACE ICE Balasubramanian and Feizi layer masking
Closest papers by meaning
- MCPNet: An Interpretable Classifier via Multi-Level Concept Prototypes (CVPR 2024 · similarity 0.85)
- I-CEE: Tailoring Explanations of Image Classification Models to User Expertise (AAAI 2024 · similarity 0.84)
- FastCAV: Efficient Computation of Concept Activation Vectors for Explaining Deep Neural Networks (ICML 2025 · similarity 0.84)
- CoE: Chain-of-Explanation via Automatic Visual Concept Circuit Description and Polysemanticity Quantification (CVPR 2025 · similarity 0.84)
- FACE: Faithful Automatic Concept Extraction (NeurIPS 2025 · similarity 0.84)
- Show and Tell: Visually Explainable Deep Neural Nets via Spatially-Aware Concept Bottleneck Models (CVPR 2025 · similarity 0.84)
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.