POEM: Pattern-Oriented Explanations of CNN Models

Download PDF

Abstract

Deep learning models achieve state-of-the-art performance in many applications, but their prediction decisions are difficult to explain. Various solutions exist in the area of explainable AI, for example to understand individual predictions or to approximate complex models using simpler interpretable ones. We contribute to this body of work with POEM: a tool that produces pattern-oriented explanations of deep learning models. POEM explains models that learn hierarchies of concepts, such as Convolutional Neural Networks that detect various shapes and objects in images. For example, POEM may identify a pattern of the form ``if an image contains a bed then the given model classifies the image as a bedroom rather than a kitchen''. We present the modular design of POEM, followed by examples of POEM's use in model auditing and detecting errors in training data.