Explaining A Black-box By Using A Deep Variational Information Bottleneck Approach

Interpretable machine learning has gained much attention recently. Briefness and comprehensiveness are necessary in order to provide a large amount of information concisely when explaining black-box decision system. However, existing interpretable methods fail consider briefness simultaneously, leading redundant explanations. We propose the variational bottleneck for interpretation, VIBI, system-agnostic method that provides brief but comprehensive explanation. VIBI adopts an theoretic principle, as criterion finding such For each instance, selects key features maximally compressed about input (briefness), informative made by system on (comprehensive). evaluate three datasets compare with state-of-the-art terms both interpretability fidelity evaluated human quantitative metrics.