Thinking like a machine — generating visual rationales through latent space optimization

Interpretability and small labelled datasets are key issues in the practical application of deep learning, particularly in areas such as medicine. In this paper, we present a semi-supervised technique that addresses both these issues simultaneously. We learn dense representations from large unlabelled image datasets, then use those representations to both learn classifiers from small labeled sets and generate visual rationales explaining the predictions. Using chest radiography diagnosis as a motivating application, we show our method has good generalization ability by learning to represent our chest radiography dataset while training a classifier on an separate set from a different institution. Our method identifies heart failure and other thoracic diseases. For each prediction, we generate visual rationales for positive classifications by optimizing a latent representation to minimize the probability of disease while constrained by a similarity measure in image space. Decoding the resultant latent representation produces an image without apparent disease. The difference between the original and the altered image forms an interpretable visual rationale for the algorithm's prediction. Our method simultaneously produces visual rationales that compare favourably to previous techniques and a classifier that outperforms the current state-of-the-art.