2015/08/14 by Salman Khan, Khan, Salman H., Munawar Hayat +7 · 8 citations
Computer Science · Medicine · Health Professions · #Imbalanced Data Classification Techniques #Retinal Imaging and Analysis #Artificial Intelligence in Healthcare
paper · pdf · doi:10.48550/arxiv.1508.03422
Class imbalance is a common problem in the case of real-world object\ndetection and classification tasks. Data of some classes is abundant making\nthem an over-represented majority, and data of other classes is scarce, making\nthem an under-represented minority. This imbalance makes it challenging for a\nclassifier to appropriately learn the discriminating boundaries of the majority\nand minority classes. In this work, we propose a cost sensitive deep neural\nnetwork which can automatically learn robust feature representations for both\nthe majority and minority classes. During training, our learning procedure\njointly optimizes the class dependent costs and the neural network parameters.\nThe proposed approach is applicable to both binary and multi-class problems\nwithout any modification. Moreover, as opposed to data level approaches, we do\nnot alter the original data distribution which results in a lower computational\ncost during the training process. We report the results of our experiments on\nsix major image classification datasets and show that the proposed approach\nsignificantly outperforms the baseline algorithms. Comparisons with popular\ndata sampling techniques and cost sensitive classifiers demonstrate the\nsuperior performance of our proposed method.\n