Slowing Down the Weight Norm Increase in Momentum-based Optimizers

Normalization techniques, such as batch normalization (BN), have led to significant improvements in deep neural network performances. Prior studies have analyzed the benefits of the resulting scale invariance of the weights for the gradient descent (GD) optimizers: it leads to a stabilized training due to the auto-tuning of step sizes. However, we show that, combined with the momentum-based algorithms, the scale invariance tends to induce an excessive growth of the weight norms. This in turn overly suppresses the effective step sizes during training, potentially leading to sub-optimal performances in deep neural networks. We analyze this phenomenon both theoretically and empirically. We propose a simple and effective solution: at each iteration of momentum-based GD optimizers (e.g. SGD or Adam) applied on scale-invariant weights (e.g. Conv weights preceding a BN layer), we remove the radial component (i.e. parallel to the weight vector) from the update vector. Intuitively, this operation prevents the unnecessary update along the radial direction that only increases the weight norm without contributing to the loss minimization. We verify that the modified optimizers SGDP and AdamP successfully regularize the norm growth and improve the performance of a broad set of models. Our experiments cover tasks including image classification and retrieval, object detection, robustness benchmarks, and audio classification. Source code is available at https://github.com/clovaai/AdamP.