On the Ideal Number of Groups for Isometric Gradient Propagation
This addresses a hyperparameter tuning bottleneck for researchers and practitioners using group normalization, though it is incremental as it builds on existing normalization techniques.
The paper tackles the problem of determining the optimal number of groups for group normalization in deep neural networks, proposing a theoretically grounded method that sets this number based on gradient behavior to improve training. The result shows improved performance across various architectures, tasks, and datasets.
Recently, various normalization layers have been proposed to stabilize the training of deep neural networks. Among them, group normalization is a generalization of layer normalization and instance normalization by allowing a degree of freedom in the number of groups it uses. However, to determine the optimal number of groups, trial-and-error-based hyperparameter tuning is required, and such experiments are time-consuming. In this study, we discuss a reasonable method for setting the number of groups. First, we find that the number of groups influences the gradient behavior of the group normalization layer. Based on this observation, we derive the ideal number of groups, which calibrates the gradient scale to facilitate gradient descent optimization. Our proposed number of groups is theoretically grounded, architecture-aware, and can provide a proper value in a layer-wise manner for all layers. The proposed method exhibited improved performance over existing methods in numerous neural network architectures, tasks, and datasets.