Neural Collapse with Normalized Features: A Geometric Analysis over the Riemannian Manifold
This work provides a geometric analysis for a common practice in representation learning, offering theoretical insights into training dynamics, but it is incremental as it extends existing neural collapse theory to normalized features.
The paper theoretically justifies the neural collapse phenomenon for normalized features in overparameterized deep networks, showing that the only global minimizers are neural collapse solutions and all other critical points are strict saddles, with experimental results indicating faster learning of better representations.
When training overparameterized deep networks for classification tasks, it has been widely observed that the learned features exhibit a so-called "neural collapse" phenomenon. More specifically, for the output features of the penultimate layer, for each class the within-class features converge to their means, and the means of different classes exhibit a certain tight frame structure, which is also aligned with the last layer's classifier. As feature normalization in the last layer becomes a common practice in modern representation learning, in this work we theoretically justify the neural collapse phenomenon for normalized features. Based on an unconstrained feature model, we simplify the empirical loss function in a multi-class classification task into a nonconvex optimization problem over the Riemannian manifold by constraining all features and classifiers over the sphere. In this context, we analyze the nonconvex landscape of the Riemannian optimization problem over the product of spheres, showing a benign global landscape in the sense that the only global minimizers are the neural collapse solutions while all other critical points are strict saddles with negative curvature. Experimental results on practical deep networks corroborate our theory and demonstrate that better representations can be learned faster via feature normalization.