Research map / Label-efficient and robust learning / Recognition and long-tail
CHiQPM: Calibrated Hierarchical Interpretable Image 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 CHiQPM, a globally interpretable image classifier that builds a hierarchical, feature-sharing class tree enabling built-in interpretable conformal prediction with competitive set sizes and state-of-the-art compact accuracy.
Problem
Globally interpretable image classifiers with compact class representations lack comprehensive local explanations and interpretable calibrated set prediction; conformal prediction sets typically contain incoherent classes that misalign with human reasoning.
Approach
Extends QPM with a hierarchical constraint forcing many class pairs to share n-1 of their n features, a Feature Grounding Loss with ReLU that yields sparser grounded features, and a nonconformity score that traverses the resulting class hierarchy so conformal prediction sets are coherent and calibrated with split CP.
Key results
- CHiQPM reaches 85.3% accuracy on CUB with ResNet50 (dense 86.6%), and 75.3% on ImageNet, the best among compact models.
- Contrastiveness on CUB is 99.9% versus 96.0% for the strongest prior compact baseline (Q-SENN/QPM family), with structural grounding 75.0 vs 47.9.
- Calibrated built-in set predictor gives average set size 2.94 at alpha=0.075 on CUB, versus 1.67 for THR and 8.54 for APS, making it competitive with non-interpretable CP methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CHiQPM instead of NewlayerName | compact ResNet50 classifier · CUB | accuracy: 85.1 -> 85.3 | Helped |
| Use CHiQPM instead of NewlayerName | compact ResNet50 classifier · CARS | accuracy: 91.8 -> 91.9 | Helped |
| Use CHiQPM instead of Q-SENN | compact ResNet50 classifier · ImageNet | accuracy: 74.3 -> 75.3 | Helped |
| Add Feature Grounding Loss (ablation) | CHiQPM · CUB | Feature Alignment: 3.0 -> 3.8 | Helped |
| Add Feature Grounding Loss (ablation) | CHiQPM · CUB | Feature Sparsity: 39.9 -> 22.3 | Helped |
| Add Feature Grounding Loss (ablation) | CHiQPM · CUB | accuracy: 85.3 -> 85.3 | No clear effect |
| Replace Feature Grounding Loss with L1 regularization (ablation) | CHiQPM · CUB | Feature Alignment: 3.8 -> 3.4 | Hurt |
| Replace Feature Grounding Loss with L1 regularization (ablation) | CHiQPM · CUB | Feature Sparsity: 22.3 -> 31.3 | Hurt |
| Add Feature Grounding Loss and ReLU (ablation) | CHiQPM · CUB | accuracy: 84.8 -> 85.3 | Helped |
| Add Feature Grounding Loss and ReLU (ablation) | CHiQPM · CUB | Feature Alignment: 1.9 -> 3.8 | Helped |
| Add normalization (ablation) | CHiQPM · CUB | accuracy: 83.9 -> 85.3 | Helped |
| Increase Feature Diversity Loss weighting (ablation) | CHiQPM · CUB | loc5: 88.1 -> 92.7 -> 96.5 | Helped |
Surprising findings
- Calibrated built-in interpretable set prediction is less efficient than THR at high coverage, whereas it approaches APS-level efficiency only at lower alpha in some settings.
- Increasing the density parameter rho beyond 0.5 slightly deteriorates point-prediction accuracy.
Limitations
- Set predictor is less efficient than THR in several settings, and the hierarchy's efficiency depends on rho tuning on calibration data.
- Point accuracy slightly declines for rho beyond 0.5, creating a tradeoff between point and set prediction.
- Code is not released at submission; reproducibility depends on the planned release after acceptance.
Experimental setup
Datasets: CUB Stanford Cars ImageNet
Benchmarks: CUB Stanford Cars ImageNet
Metrics: accuracy compactness contrastiveness structural grounding feature alignment set size set coherence local dimensional coverage
Baselines: QPM Q-SENN SLDD-Model PIP-Net ProtoPool glm-saga THR APS
Builds on: QPM Q-SENN split conformal prediction SENN GLM-SAGA
Closest papers by meaning
- QPM: Discrete Optimization for Globally Interpretable Image Classification (ICLR 2025 · similarity 0.86)
- Enhancing Performance of Explainable AI Models with Constrained Concept Refinement (ICML 2025 · similarity 0.83)
- SLICE: Stabilized LIME for Consistent Explanations for Image Classification (CVPR 2024 · similarity 0.83)
- A Multimodal Automated Interpretability Agent (ICML 2024 · similarity 0.83)
- Q-SENN: Quantized Self-Explaining Neural Networks (AAAI 2024 · similarity 0.82)
- Everything, Everywhere, All at Once: Is Mechanistic Interpretability Identifiable? (ICLR 2025 · similarity 0.82)
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.