Assessing the usefulness of NetworkDissection in identifying the interpretability of facial characterization networks

Abstract

In this thesis I test for the emergence of policy relevant features as disentangled, human interpretable representations in facial characterization networks. I probe an age and gender classifier using the NetworkDissection method with a hand labelled dataset of facial features, skin tones, and textures. Facial features and skin tones emerge as disentangled concepts in each of the networks probed. The emergence of these features in a smaller image classification network indicates the effectiveness of the NetworkDissection method in contexts other than ones studied in the original paper. Moreover, the emergence of policy relevant features, skin tones, indicates the method may be effective in identifying policy sensitive attributes. I also analyze the robustness of the NetworkDissection technique itself to changes in a key component of the experimental setup: the source of ground truth human understandable concepts. The results demonstrate the technique reliably applies labels when new concepts and samples are added to the set of ground truth labels.