Mitigating Dataset Bias Using Per-Sample Gradients From A Biased Classifier

The performance of deep neural networks (DNNs) primarily depends on the configuration of the training set. Specifically, biased training sets can make the trained model have unintended prejudice, which causes severe errors in the inference. Although several studies have addressed biased training using human supervision, few studies have been conducted without human knowledge because biased information cannot be easily extracted without human involvement. This study proposes a simple method to remove prejudice from a biased model without additional information and reconstruct a balanced training set based on the biased training set. The novel training method consists of three steps: (1) training biased DNNs, (2) measuring the contribution to the prejudicial training and generating balanced data batches to prevent the prejudice, (3) training de-biased DNNs with the balanced data. We test the training method based on various synthetic and real-world biased sets and discuss how gradients can efficiently detect minority samples. The experiment demonstrates that the detection method based on the gradients helps erase prejudice, resulting in improved inference accuracy by up to 19.58\% compared to the other state-of-the-art algorithm.