Research map / Label-efficient and robust learning / Recognition and long-tail
Enhancing Performance of Explainable AI Models with Constrained Concept Refinement
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 Constrained Concept Refinement (CCR), which refines CLIP-based concept embeddings within a small norm ball for IP-OMP-style interpretable-by-design models, with theory on a generative test-bed and image classification experiments.
Problem
Interpretable-by-design models (IP-OMP, CBMs) lose accuracy because concept embeddings from pretrained encoders such as CLIP are inaccurate or misaligned, and existing fixes add black-box modules that undermine interpretability.
Approach
Directly optimize concept embeddings under a norm constraint around their initialization (projected gradient descent on a differentiable surrogate of IP-OMP), so the radius rho trades accuracy against interpretability; theory shows zero loss and convergence to ground-truth latent features on a sparse generative model.
Key results
- CCR outperforms CLIP-IP-OMP and lf-CBM on CIFAR-10, CIFAR-100, ImageNet and Places365 but not on CUB-200 (lf-CBM uses a CUB-specific ResNet-18).
- On ImageNet, CCR (k about 49) runs in about 2 hours versus about 33 hours for CLIP-IP-OMP at k=50 on a V100, with higher accuracy.
- Theory: CCR drives the aggregated squared loss to zero and the latent-matrix error contracts linearly under the generative model.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Refine concept embeddings with CCR instead of using the unrefined CCR baseline (ablation) | CCR image classifier · CIFAR-10, CIFAR-100, ImageNet, CUB-200, Places365 | test accuracy | Helped |
| Use CCR rather than CLIP-IP-OMP | CLIP-based explainable image classification · CIFAR-10, CIFAR-100, ImageNet, Places365 | prediction accuracy | Helped |
| Use CCR rather than lf-CBM | Explainable image classification · CIFAR-10, CIFAR-100, ImageNet, Places365 | prediction accuracy | Helped |
| Use CCR rather than lf-CBM | Explainable image classification · CUB-200 | prediction accuracy | Hurt |
| Run CLIP-IP-OMP with explanation length k=10 instead of k≈50 | CLIP-IP-OMP on ImageNet · ImageNet | accuracy | Hurt |
| Use CCR rather than CLIP-IP-OMP with k≈49 | ImageNet image classification · ImageNet | runtime: ≈ 33 hours -> ≈ 2 hours | Helped |
| Optimize concept embeddings from the initial embeddings | CCR under the paper's generative model · Generative model | squared loss: Theorem: \mathcal{L}(\Tilde{f}_{\mD^{(t+1)}})\le \left(1-2\eta\|\rvx\|^2_2\right)^2\mathcal{L}(\Tilde{f}_{\mD^{(t)}}) | Helped |
Surprising findings
- Naive CLIP-IP-OMP at k=10 drops to about 63% accuracy on ImageNet, showing strong sensitivity to explanation length.
Limitations
- Theory uses a Gaussian sparse generative test-bed with column-orthogonal, full-rank latent matrix assumptions that do not hold exactly for real images.
- Hard-thresholding replaces exact top-k so sparsity level k is only controlled implicitly through lambda.
- CUB-200 is not improved over lf-CBM, which uses a backbone trained on CUB-200.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 ImageNet CUB-200 Places365
Benchmarks: CIFAR-10 CIFAR-100 ImageNet CUB-200 Places365
Metrics: test accuracy average explanation length average sparsity ratio average concept embedding deviation runtime
Baselines: CLIP-IP-OMP label-free CBM CCR without refinement (eta_D=0)
Builds on: Information Pursuit IP-OMP label-free CBM task-driven dictionary learning
Closest papers by meaning
- Discovering Interpretable Representations for Both Deep Generative and Discriminative Models (ICML 2018 · similarity 0.88)
- Restyling Unsupervised Concept Based Interpretable Networks with Generative Models (ICLR 2025 · similarity 0.88)
- Predictive Concept Decoders: Training Scalable End-to-End Interpretability Assistants (NeurIPS 2026 · similarity 0.87)
- Interpretable Generative Models through Post-hoc Concept Bottlenecks (CVPR 2025 · similarity 0.87)
- Explainable Neural Networks with Guarantee: A Sparse Estimation Approach (AAAI 2025 · similarity 0.87)
- Enhancing Pre-trained Representation Classifiability can Boost its Interpretability (ICLR 2025 · similarity 0.87)
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.