Research map / Label-efficient and robust learning / Recognition and long-tail
Enhancing Pre-trained Representation Classifiability can Boost its Interpretability
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 Interpretability-Interpretation Score (IIS), a metric quantifying representation interpretability via concept-based sparsification, and shows that higher classifiability and interpretability of pre-trained visual representations mutually promote each other, leveraging this to enhance interpretable classification.
Problem
Pre-trained visual representations are optimized for downstream classifiability, but it is unclear whether they can also be highly interpretable; no general metric existed to quantify representation interpretability and its relation to classifiability.
Approach
Defines IIS as the area under the curve of accuracy retention ratio versus concept sparsity, where representations are projected onto a concept library and less relevant concepts are removed; empirically finds a mutually promoting relationship between interpretability and classifiability, and uses IIS maximization as an auxiliary loss to improve both.
Key results
- IIS increases monotonically with classifiability across ResNet-18 to ResNet-152 on ImageNet-1k (e.g., 0.695 to 0.780 IIS; Acc@1 63.39 to 68.00).
- On CUB-200, CLIP-ViT-B/16 reaches IIS 0.808 and Acc@1 80.6 versus ViT-B/16 at 0.764 and 75.9.
- On CUB-200, the method reaches 83.50 Acc@1 with 247 concepts versus ECBM at 83.21 under the same concept library.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CLIP-ViT-B/16 instead of ViT-B/16 | pre-trained representations · CUB-200 | IIS: 0.764 -> 0.808 | Helped |
| Use CLIP-ViT-B/16 instead of ViT-B/16 | pre-trained representations · CUB-200 | Acc@1: 75.9 -> 80.6 | Helped |
| Use Ours instead of LaBo | image classification · UCF101, DTD, HAM10000 | accuracy: 97.68 -> 97.79; 76.86 -> 77.22; 81.40 -> 81.90 | Helped |
| Use Ours instead of LF-CBM | ViT-B and ViT-L · ImageNet-1k and CUB-200 | accuracy: 77.94 -> 78.50; 68.97 -> 69.14; 85.52 -> 86.85; 82.83 -> 83.51 | Helped |
| Maximize IIS before segmentation fine-tuning | ResNet-50 representations · ADE20K | mIoU: 40.60 -> 42.35 | Helped |
| Use Ours instead of ECBM with the same 247-concept library | CUB-200 classification · CUB-200 | Acc@1: 83.21 -> 83.50 | Helped |
| Use Ours instead of DEAL | CLIP-RN50 and CLIP-ViT-B/32 · ImageNet-1k | accuracy: 70.0 -> 70.8; 70.8 -> 73.0 | Helped |
| Replace concept sparsification with clustering (ablation) | IIS computation on ResNet-series representations · CUB-200 | IIS: IIS-Cluster 0.664, 0.675, 0.721, 0.732, 0.742; IIS 0.695, 0.719, 0.766, 0.776, 0.780 | Hurt |
| Replace Equation 2 with cos-cubed loss (ablation) | IIS evaluation on ResNet-series representations · CUB-200 | IIS: 0.756, 0.774, 0.789, 0.802, 0.805 | No clear effect |
| Mask small absolute-value elements without adjusting the remaining elements (ablation) | IIS sparsification on ResNet-series representations · CUB-200 | IIS: 0.695 -> 0.769; 0.719 -> 0.774; 0.766 -> 0.809; 0.776 -> 0.816; 0.780 -> 0.828 | Helped |
| Remove concepts in descending rather than ascending similarity order (ablation) | IIS concept sparsification on ResNet-series models · CUB-200 | IIS: 0.695 -> 0.284; 0.719 -> 0.298; 0.766 -> 0.340; 0.776 -> 0.352; 0.780 -> 0.354 | Hurt |
| Increase the sparsity ratio above s=0.1 (ablation) | IIS-maximization training · Table 3 experiments | performance | Hurt |
Surprising findings
- Masking-only sparsification (no value adjustment) yields higher IIS than soft-threshold sparsification (e.g., 0.828 vs 0.780 for ResNet-152).
- Randomly initialized models can show high IIS because accuracy retention ratio is high when both interpretations and representations have low accuracy.
Limitations
- IIS is designed for classifiability-oriented pre-trained representations and does not handle extremely low-classifiability (random) representations well.
- Trustworthiness of concept-based interpretations is not evaluated in the metric.
- Most results are single runs without standard deviations; only Table 4 includes std over 5 seeds.
Experimental setup
Datasets: ImageNet-1k CUB-200 Kinetics-400 ADE20K UCF-101 DTD HAM10000
Benchmarks: ImageNet-1k CUB-200 Kinetics-400 ADE20K UCF-101 DTD HAM10000
Metrics: IIS Acc@1 mIoU
Baselines: LaBo LF-CBM ECBM DEAL VLG-CBM
Builds on: concept bottleneck models label-free concept bottleneck models CLIP
Closest papers by meaning
- Measuring Per-Unit Interpretability at Scale Without Humans (NeurIPS 2024 · similarity 0.87)
- Enhancing Performance of Explainable AI Models with Constrained Concept Refinement (ICML 2025 · similarity 0.87)
- Restyling Unsupervised Concept Based Interpretable Networks with Generative Models (ICLR 2025 · similarity 0.86)
- Towards Fine-Grained Interpretability: Counterfactual Explanations for Misclassification with Saliency Partition (CVPR 2025 · similarity 0.86)
- Perceptual Inductive Bias Is What You Need Before Contrastive Learning (CVPR 2025 · similarity 0.86)
- Causal Tracing of Object Representations in Large Vision Language Models: Mechanistic Interpretability and Hallucination Mitigation (AAAI 2026 · 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.