Spherical Perspective on Learning with Normalization Layers
This provides incremental insights into optimization dynamics for researchers in deep learning, but does not address a broad practical problem.
The paper tackled the problem of understanding the optimization effects of normalization layers in deep learning by introducing a spherical geometric framework, resulting in the derivation of Adam's first effective learning rate expression and showing that SGD with normalization layers is equivalent to a constrained variant of Adam.
Normalization Layers (NLs) are widely used in modern deep-learning architectures. Despite their apparent simplicity, their effect on optimization is not yet fully understood. This paper introduces a spherical framework to study the optimization of neural networks with NLs from a geometric perspective. Concretely, the radial invariance of groups of parameters, such as filters for convolutional neural networks, allows to translate the optimization steps on the $L_2$ unit hypersphere. This formulation and the associated geometric interpretation shed new light on the training dynamics. Firstly, the first effective learning rate expression of Adam is derived. Then the demonstration that, in the presence of NLs, performing Stochastic Gradient Descent (SGD) alone is actually equivalent to a variant of Adam constrained to the unit hypersphere, stems from the framework. Finally, this analysis outlines phenomena that previous variants of Adam act on and their importance in the optimization process are experimentally validated.