Cost Sensitive Learning of Deep Feature Representations from Imbalanced Data

In real-world object detection and classification tasks, data for common classes appear quite frequently (majority) while rarer classes have a lower representation (minority). With the imbalanced data, it therefore becomes challenging for a classifier to learn equally good boundaries for the majority and minority classes. In this work, we propose a cost sensitive deep neural network which can automatically learn robust feature representations for both the majority and minority classes. During training, our learning procedure involves an alternative joint optimization for the class dependent costs and the neural network parameters. The proposed approach is applicable to both binary and multi-class problems without any modification. Moreover, it does not add to the computational load during training, because the original data distribution is not disturbed. We report experiments on four major datasets relating to image classification and show that the proposed approach significantly outperforms the baseline procedures. The comparisons with the popular data sampling techniques and the cost sensitive classifiers demonstrate superior performance of our proposed method.