Concept-Based Explanations

Concept-based explanations study how to describe a model’s decisions using meaningful features—such as objects, clinical findings, or sound patterns—instead of only pixels or hidden activations. Approaches include concept bottlenecks and activation vectors, sparse autoencoders and dictionaries, language-grounded concepts, and concept graphs or circuits; they are used in settings from image classification and medical diagnosis to time series and language models. Researchers test explanations with concept insertion, deletion, ablation, counterfactuals, and comparisons with human judgments, while also using concepts to audit shortcuts or steer model behavior. A central challenge is ensuring that concepts are both understandable and faithful: a feature may be easy to identify without actually driving the prediction, and greater transparency can sometimes come with reduced accuracy or incomplete explanations.