Convolutional Neural Networks (CNNs) are commonly used in computer vision. However, their predictions are difficult to explain, as is the case with many deep learning models. To address this problem, we present POEM, a modular framework that produces patterns of semantic concepts such as shapes and colours to explain image classifier CNNs. POEM identifies patterns such as "if sofa then living room", meaning that if an image contains a sofa and the model pays attention to the sofa, then the model classifies the image as a living room. We illustrate the advantages of POEM over existing work using quantitative and qualitative experiments.

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 image classification models. POEM explains models that learn hierarchies of concepts, such as Convolutional Neural Networks that detect shapes and objects in images. For example, POEM may identify a pattern of the form "if bed then bedroom", indicating that if an image contains a bed and the model pays attention to this region of the image during inference, then the model classifies the image as a bedroom. We present the modular design of POEM, followed by examples of POEM's use in model auditing and detecting errors in training data.