Dynamics of stochastic gradient descent for two-layer neural networks in the teacher–student setup - Sorbonne Université
Journal Articles Journal of Statistical Mechanics: Theory and Experiment Year : 2020

Dynamics of stochastic gradient descent for two-layer neural networks in the teacher–student setup

Abstract

Deep neural networks achieve stellar generalisation even when they have enough parameters to easily fit all their training data. We study this phenomenon by analysing the dynamics and the performance of over-parameterised two-layer neural networks in the teacher–student setup, where one network, the student, is trained on data generated by another network, called the teacher. We show how the dynamics of stochastic gradient descent (SGD) is captured by a set of differential equations and prove that this description is asymptotically exact in the limit of large inputs. Using this framework, we calculate the final generalisation error of student networks that have more parameters than their teachers. We find that the final generalisation error of the student increases with network size when training only the first layer, but stays constant or even decreases with size when training both layers. We show that these different behaviours have their root in the different solutions SGD finds for different activation functions. Our results indicate that achieving good generalisation in neural networks goes beyond the properties of SGD alone and depends on the interplay of at least the algorithm, the model architecture, and the data set.
Fichier principal
Vignette du fichier
Goldt_2020_J._Stat._Mech._2020_124010.pdf (1.01 Mo) Télécharger le fichier
Origin Publication funded by an institution

Dates and versions

hal-03289916 , version 1 (19-07-2021)

Identifiers

Cite

Sebastian Goldt, Madhu S Advani, Andrew M Saxe, Florent Krzakala, Lenka Zdeborová. Dynamics of stochastic gradient descent for two-layer neural networks in the teacher–student setup. Journal of Statistical Mechanics: Theory and Experiment, 2020, 2020 (12), pp.124010. ⟨10.1088/1742-5468/abc61e⟩. ⟨hal-03289916⟩
35 View
58 Download

Altmetric

Share

More