Naming the Concepts Classifiers Rely On: Language-Anchored Decomposition for Faithful Explanation

2026-07-08

AI Summary

Existing model explanation methods struggle to simultaneously achieve concept interpretability, faithfulness to model behavior, and non-intrusiveness. This work proposes the Language-Anchored Decomposition (LAD) framework, which for the first time unifies these three desiderata in post-hoc explanations. LAD leverages a large language model to generate human-readable concept names as structural priors, employs CLIP-based similarity maps to localize relevant image regions, and—without modifying the original frozen model—learns only a concept basis that reconstructs encoder activations using language-anchored maps as coefficient matrices. By integrating vision-language alignment with non-negative matrix decomposition, LAD produces spatially precise, decision-relevant explanations across natural images, scenes, and medical imaging, demonstrates high faithfulness in concept insertion and deletion tests, and yields stable, interpretable concept names.

Abstract

Deep neural networks are widely deployed in high-stakes visual applications where interpretability is critical, yet existing explanations face a trade-off: post-hoc concept methods recover factors that are faithful to a model's behavior but unnamed, while naming and by-design methods attach human-readable concepts only by retraining or altering the classifier. We propose Language-Anchored Decomposition (LAD), a post-hoc framework that delivers concepts which are simultaneously named, faithful, and obtained without modifying the model. For each class, a large language model proposes a concept vocabulary that CLIP-based similarity maps localize across image regions. Inverting standard non-negative matrix factorization, LAD fixes these language-grounded maps as the coefficient matrix and learns only a concept basis that reconstructs the frozen encoder's activations, so naming becomes a structural constraint and the model's own feature geometry determines which concepts are retained. Removing this anchor preserves accuracy but collapses attribution faithfulness. Across natural-image, scene, and medical-imaging benchmarks, LAD produces spatially precise explanations that are decision-relevant under both concept insertion and deletion, while uniquely providing stable, human-interpretable concept names.

Problem

• interpretability
• post-hoc explanation
• concept naming
• faithful explanation
• visual classifiers

Innovation

• Language-Anchored Decomposition
• post-hoc explanation
• concept-based interpretability
• CLIP
• non-negative matrix factorization