Provable concept learning for interpretable predictions using variational autoencoders

In safety-critical applications, practitioners are reluctant to trust neural\nnetworks when no interpretable explanations are available. Many attempts to\nprovide such explanations revolve around pixel-based attributions or use\npreviously known concepts. In this paper we aim to provide explanations by\nprovably identifying \\emph{high-level, previously unknown ground-truth\nconcepts}. To this end, we propose a probabilistic modeling framework to derive\n(C)oncept (L)earning and (P)rediction (CLAP) -- a VAE-based classifier that\nuses visually interpretable concepts as predictors for a simple classifier.\nAssuming a generative model for the ground-truth concepts, we prove that CLAP\nis able to identify them while attaining optimal classification accuracy. Our\nexperiments on synthetic datasets verify that CLAP identifies distinct\nground-truth concepts on synthetic datasets and yields promising results on the\nmedical Chest X-Ray dataset.\n