Entropic gradient descent algorithms and wide flat minima*This article is an updated version of: Pittorino F, Lucibello C, Feinauer C, Perugini G, Baldassi C, Demyanenko E and Zecchina R 2021 Entropic gradient descent algorithms and wide flat minima Int. Conf. on Learning Representations. (29th December 2021)

• Record Type: Journal Article
• Authors: Pittorino, Fabrizio; Lucibello, Carlo; Feinauer, Christoph; Perugini, Gabriele; Baldassi, Carlo; Demyanenko, Elizaveta; Zecchina, Riccardo
• Is Part Of: Journal of statistical mechanics. (2021:Dec.)
• Publication Date: 2021-12-29
• Subjects: deep learning -- machine learning -- message-passing algorithms; Statistical mechanics -- Periodicals; Mechanics -- Statistical methods -- Periodicals; 530.1305
• DOI: 10.1088/1742-5468/ac3ae8
• Languages: English
• ISSNs: 1742-5468

Abstract

Abstract: The properties of flat minima in the empirical risk landscape of neural networks have been debated for some time. Increasing evidence suggests they possess better generalization capabilities with respect to sharp ones. In this work we first discuss the relationship between alternative measures of flatness: the local entropy, which is useful for analysis and algorithm development, and the local energy, which is easier to compute and was shown empirically in extensive tests on state-of-the-art networks to be the best predictor of generalization capabilities. We show semi-analytically in simple controlled scenarios that these two measures correlate strongly with each other and with generalization. Then, we extend the analysis to the deep learning scenario by extensive numerical validations. We study two algorithms, entropy-stochastic gradient descent and replicated-stochastic gradient descent, that explicitly include the local entropy in the optimization objective. We devise a training schedule by which we consistently find flatter minima (using both flatness measures), and improve the generalization error for common architectures (e.g. ResNet, EfficientNet).