Towards Explainable Image Classifier: An Analogy to Multiple Choice Question Using Patch-level Similarity Measure

Abstract

With increased interests in Explainable Artificial Intelligence (XAI), many researches find ways to provide explanations for machine learning algorithms and their predictions. We propose Multiple Choice Questioned Convolutional Neural Network (MCQ-CNN) to better understand the prediction of image classifier by considering the problem of multi-class classification as the problem of multiple choice question. MCQ-CNN not only performs classification of the query image, but also explains the classification result by demonstrating the elimination process of multi-class labels in patch-level. The proposed model consists of two modules. Classification module is to classify class label of the query. Elimination module is to perform similarity measure in patch-level to distinguish whether the target object part shares the feature of certain class label or not. Classification module is trained using ResNet with high classification accuracy. Elimination module performs similarity measure by distance metric learning based on Large Margin Nearest Neighbor (LMNN). Our experiments have shown notable performances in both classification and elimination modules.