Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV)

TLDR

Deep‑learning models are difficult to interpret because of their size, complexity, and reliance on low‑level features rather than high‑level concepts. The authors propose Concept Activation Vectors (CAVs) to interpret a neural network’s internal state using human‑friendly concepts. CAVs are used in Testing with CAVs (TCAV), which applies directional derivatives to quantify how much a user‑defined concept (e.g., stripes for zebra) influences a classification, enabling hypothesis testing in image and medical domains.

Abstract

The interpretation of deep learning models is a challenge due to their size, complexity, and often opaque internal state. In addition, many systems, such as image classifiers, operate on low-level features rather than high-level concepts. To address these challenges, we introduce Concept Activation Vectors (CAVs), which provide an interpretation of a neural net's internal state in terms of human-friendly concepts. The key idea is to view the high-dimensional internal state of a neural net as an aid, not an obstacle. We show how to use CAVs as part of a technique, Testing with CAVs (TCAV), that uses directional derivatives to quantify the degree to which a user-defined concept is important to a classification result--for example, how sensitive a prediction of "zebra" is to the presence of stripes. Using the domain of image classification as a testing ground, we describe how CAVs may be used to explore hypotheses and generate insights for a standard image classification network as well as a medical application.