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

2026-07-08 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

Language-Anchored Decomposition (LAD) is a novel post-hoc framework designed to provide named, faithful, and model-agnostic explanations for deep neural networks in high-stakes visual applications. Unlike existing methods that trade off faithfulness for human-readable concepts or require model retraining, LAD achieves both without modifying the original classifier. It operates by having a large language model propose a concept vocabulary for each class, which CLIP-based similarity then localizes across image regions. LAD inverts standard non-negative matrix factorization, fixing these language-grounded maps as the coefficient matrix and learning a concept basis that reconstructs the frozen encoder's activations. This process structurally constrains naming, allowing the model's feature geometry to determine concept retention. Removing this language anchor preserves accuracy but compromises attribution faithfulness. LAD produces spatially precise, decision-relevant explanations across natural-image, scene, and medical-imaging benchmarks, offering stable, human-interpretable concept names.

Key takeaway

For AI Scientists and Machine Learning Engineers deploying deep neural networks in critical visual applications, Language-Anchored Decomposition (LAD) offers a robust solution for model interpretability. You can now generate faithful, human-interpretable concept explanations post-hoc, without the need to retrain or alter your existing models. This capability is crucial for validating model decisions and building trust, especially in domains like medical imaging where transparency is paramount. Consider integrating LAD to enhance the explainability of your deployed vision systems.

Key insights

LAD provides named, faithful, post-hoc explanations for deep neural networks without modifying the original model.

Principles

• Language models can anchor concept vocabularies.
• Inverted NMF can structurally constrain naming.
• Model's feature geometry determines concept retention.

Method

LAD uses a large language model for concept vocabulary, CLIP for localization, then inverts non-negative matrix factorization to learn a concept basis from frozen encoder activations, ensuring named and faithful explanations.

In practice

• Explain high-stakes visual AI decisions.
• Generate spatially precise concept maps.
• Benchmark explanation faithfulness with insertion/deletion.

Topics

• Language-Anchored Decomposition
• Explainable AI
• Deep Neural Networks
• Post-hoc Interpretability
• CLIP
• Non-negative Matrix Factorization
• Medical Imaging

Best for: Computer Vision Engineer, AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Towards Intrinsic Interpretability of Large Language Models:A Survey of Design Principles and Architectures — Intrinsic interpretability builds transparency directly into LLM architectures to enhance trustworthiness and safety.
• Interview with Deepika Vemuri: interpretability and concept-based learning — Concept-based learning enhances AI interpretability by aligning model representations with human-understandable concepts.
• Applied Explainability for Large Language Models: A Comparative Study — Gradient-based explainability provides stable LLM insights, while attention is efficient but less aligned with predictions.
• [Linkpost] Interpreting Language Model Parameters — VPD decomposes language model parameters into causally important subcomponents, improving mechanistic interpretability and circuit analysis.
• Interpreting Brain Responses to Language with Sparse Features from Language Models — Augmented Sparse Encoding Models reveal interpretable, general LM features align with human brain language processing.
• H-Probes: Extracting Hierarchical Structures From Latent Representations of Language Models — LLMs geometrically represent hierarchical structures like tree depth and pairwise distance in low-dimensional latent subspaces.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.