Gradient Weighted Loss for Learning from Imbalanced Samples

Training Deep Neural Networks (DNNs) with imbalanced data is indeed a daunting task. Categories with limited training data often suffer from poor representation, resulting in reduced recognition accuracy. To tackle this issue, we introduced a gradient-weighted loss function when training the deep model. Our proposed loss function leverages the average gradients across each category in the validation set to determine the importance of training data for updating the DNN's parameters. Categories with larger gradients are taken as more important than the smaller ones. In the meanwhile, our proposed loss function can combine with existing loss functions and improve model performance further on the imbalanced training dataset. Our experiments, conducted on the imbalanced CIFAR-10 and CIFAR-100 datasets, demonstrate that our proposed approach outperforms the compared methods.