Rethinking Gauss-Newton for learning over-parameterized models
This work addresses optimization challenges for machine learning practitioners by analyzing GN's convergence and generalization trade-offs in over-parameterized models, though it is incremental as it builds on existing GN and mean-field theory.
The paper tackles the optimization of over-parameterized neural networks using Gauss-Newton (GN) by establishing global convergence with faster rates than gradient descent (GD) and empirically showing a trade-off where GN finds global optima quickly but generalizes well only with specific initialization and step sizes, recovering features despite sub-optimal performance.
This work studies the global convergence and implicit bias of Gauss Newton's (GN) when optimizing over-parameterized one-hidden layer networks in the mean-field regime. We first establish a global convergence result for GN in the continuous-time limit exhibiting a faster convergence rate compared to GD due to improved conditioning. We then perform an empirical study on a synthetic regression task to investigate the implicit bias of GN's method. While GN is consistently faster than GD in finding a global optimum, the learned model generalizes well on test data when starting from random initial weights with a small variance and using a small step size to slow down convergence. Specifically, our study shows that such a setting results in a hidden learning phenomenon, where the dynamics are able to recover features with good generalization properties despite the model having sub-optimal training and test performances due to an under-optimized linear layer. This study exhibits a trade-off between the convergence speed of GN and the generalization ability of the learned solution.