Analyzing the Relations of Misclassified Inputs Between Models

Abstract

A common belief in the machine learning community is that many of the misclassified images are “difficult” images (e.g., the differentiation between classes is based on small details). We compare the misclassified images of various deep learning models and check which model misclassifies which image. We find that the misclassified images of each model are different. Moreover, despite having similar accuracy on ImageNet, one model can classify correctly more than 15% of the misclassified images of another model. This can encourage further research to use two or more architectures when performing a prediction, such as ensemble methods.