Understanding and Scheduling Weight Decay

Weight decay is a popular and even necessary regularization technique for training deep neural networks that generalize well. Previous work usually interpreted weight decay as a Gaussian prior from the Bayesian perspective. However, weight decay sometimes shows mysterious behaviors beyond the conventional understanding. For example, the optimal weight decay value tends to be zero given long enough training time. Moreover, existing work typically failed to recognize the importance of scheduling weight decay during training. Our work aims at theoretically understanding novel behaviors of weight decay and designing schedulers for weight decay in deep learning. This paper mainly has three contributions. First, we propose a novel theoretical interpretation of weight decay from the perspective of learning dynamics. Second, we propose a novel weight-decay linear scaling rule for large-batch training that proportionally increases weight decay rather than the learning rate as the batch size increases. Third, we provide an effective learning-rate-aware scheduler for weight decay, called the Stable Weight Decay (SWD) method, which, to the best of our knowledge, is the first practical design for weight decay scheduling. In our various experiments, the SWD method often makes improvements over $L_{2}$ Regularization and Decoupled Weight Decay.