Attentional-Biased Stochastic Gradient Descent

In this paper, we present a simple yet effective method (ABSGD) for addressing the data imbalance issue in deep learning. Our method is a simple modification to momentum SGD where we leverage an attentional mechanism to assign an individual importance weight to each gradient in the mini-batch. Unlike many existing heuristic-driven methods for tackling data imbalance, our method is grounded in {\it theoretically justified distributionally robust optimization (DRO)}, which is guaranteed to converge to a stationary point of an information-regularized DRO problem. The individual-level weight of a sampled data is systematically proportional to the exponential of a scaled loss value of the data, where the scaling factor is interpreted as the regularization parameter in the framework of information-regularized DRO. Compared with existing class-level weighting schemes, our method can capture the diversity between individual examples within each class. Compared with existing individual-level weighting methods using meta-learning that require three backward propagations for computing mini-batch stochastic gradients, our method is more efficient with only one backward propagation at each iteration as in standard deep learning methods. To balance between the learning of feature extraction layers and the learning of the classifier layer, we employ a two-stage method that uses SGD for pretraining followed by ABSGD for learning a robust classifier and finetuning lower layers. Our empirical studies on several benchmark datasets demonstrate the effectiveness of the proposed method.