Understanding CNN fragility when learning with imbalanced data

Abstract Convolutional neural networks (CNNs) have achieved impressive results on imbalanced image data, but they still difficulty generalizing to minority classes and their decisions are difficult interpret. These problems related because the method by which CNNs generalize classes, requires improvement, is wrapped in a black-box. To demystify CNN we focus latent features. Although embed pattern knowledge learned from training set model parameters, effect of this contained feature classification embeddings ( FE CE ). can be extracted trained global, class properties (e.g., frequency, magnitude identity) analyzed. We find that important information regarding ability network resides top-K . show learns limited number per category, magnitudes vary based whether same balanced or imbalanced. hypothesize diversity as examples, has implications for re-sampling cost-sensitive methods. methods generally rebalancing weights, numbers margins; instead diversifying also demonstrate test data if its features do not match set. use three popular datasets two algorithms commonly employed learning our experiments.