Mean-Field Langevin Dynamics and Energy Landscape of Neural Networks

TL;DR

This paper investigates the theoretical convergence of Mean-Field Langevin Dynamics in neural network training, demonstrating exponential convergence to a unique energy minimizer in the space of probability measures without restrictive assumptions.

Contribution

It introduces a novel convergence proof for MFLD using a generalized LaSalle's invariance principle and HWI inequality, applicable to general convex objectives and non-symmetric interactions.

Findings

Proves exponential convergence of MFLD to a stationary distribution.

Shows the error between finite and infinite-dimensional problems decreases as 1 over the number of parameters.

Establishes convergence without assuming convolution or symmetry in interaction potentials.

Abstract

Our work is motivated by a desire to study the theoretical underpinning for the convergence of stochastic gradient type algorithms widely used for non-convex learning tasks such as training of neural networks. The key insight, already observed in the works of Mei, Montanari and Nguyen (2018), Chizat and Bach (2018) as well as Rotskoff and Vanden-Eijnden (2018), is that a certain class of the finite-dimensional non-convex problems becomes convex when lifted to infinite-dimensional space of measures. We leverage this observation and show that the corresponding energy functional defined on the space of probability measures has a unique minimiser which can be characterised by a first-order condition using the notion of linear functional derivative. Next, we study the corresponding gradient flow structure in 2-Wasserstein metric, which we call Mean-Field Langevin Dynamics (MFLD), and show that the flow of marginal laws induced by the gradient flow converges to a stationary distribution, which is exactly the minimiser of the energy functional. We observe that this convergence is exponential under conditions that are satisfied for highly regularised learning tasks. Our proof of convergence to stationary probability measure is novel and it relies on a generalisation of LaSalle's invariance principle combined with HWI inequality. Importantly, we assume neither that interaction potential of MFLD is of convolution type nor that it has any particular symmetric structure. Furthermore, we allow for the general convex objective function, unlike, most papers in the literature that focus on quadratic loss. Finally, we show that the error between finite-dimensional optimisation problem and its infinite-dimensional limit is of order one over the number of parameters.