Tcav: Relative Concept Importance Testing with Linear Concept Activation Vectors

Neural networks commonly offer high utility but remain difficult to interpret. Developing methods to explain their decisions is challenging due to their large size, complex structure, and inscrutable internal representations. This work argues that the language of explanations should be expanded from that of input features (e.g., assigning importance weightings to pixels) to include that of higher-level, humanfriendly concepts. For example, an understandable explanation of why an image classifier outputs the label “zebra” would ideally relate to concepts such as “stripes” rather than a set of particular pixel values. This paper introduces the “concept activation vector” (CAV) which allows quantitative analysis of a concept’s relative importance to classification, with a user-provided set of input data examples defining the concept. CAVs may be easily used by non-experts, who need only provide examples, and with CAVs the high-dimensional structure of neural networks turns into an aid to interpretation, rather than an obstacle. Using the domain of image classification as a testing ground, we describe how CAVs may be used to test hypotheses about classifiers and also generate insights into the deficiencies and correlations in training data. CAVs also provide us a directed approach to choose the combinations of neurons to visualize with the DeepDream technique, which traditionally has chosen neurons or linear combinations of neurons at random to visualize.