Patchnet: Interpretable Neural Networks for Image Classification

Understanding how a complex machine learning model makes a classification decision is essential for its acceptance in sensitive areas such as health care. In certain instances, using features derived from the entire image (i.e., the global image context) might cause a model to make a correct classification decision for incorrect reasons. To resolve this problem, we present PatchNet, a method that provides a tradeoff between restricting global image context and classification error. We mathematically analyze this tradeoff using PatchNet, demonstrate how our method allows us to construct sharp visual heatmap representations of the learned features, and quantitatively compare how these features align with features selected by domain experts by applying PatchNet for the classification of cracked versus perforated textures from the Describable Textures Dataset (DTD), mouse versus human cell nuclei, normal versus fibrocystic cell nuclei, and benign versus malignant skin lesions from the ISBI-ISIC 2017 melanoma classification challenge.