Salvato in:
Dettagli Bibliografici
Autori principali: Zhu, Tongtian, He, Fengxiang, Zhang, Lan, Niu, Zhengyang, Song, Mingli, Tao, Dacheng
Natura: Preprint
Pubblicazione: 2022
Soggetti:
Accesso online:https://arxiv.org/abs/2206.12680
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866909576974041088
author Zhu, Tongtian
He, Fengxiang
Zhang, Lan
Niu, Zhengyang
Song, Mingli
Tao, Dacheng
author_facet Zhu, Tongtian
He, Fengxiang
Zhang, Lan
Niu, Zhengyang
Song, Mingli
Tao, Dacheng
contents This paper studies the algorithmic stability and generalizability of decentralized stochastic gradient descent (D-SGD). We prove that the consensus model learned by D-SGD is $\mathcal{O}{(N^{-1}+m^{-1} +λ^2)}$-stable in expectation in the non-convex non-smooth setting, where $N$ is the total sample size, $m$ is the worker number, and $1+λ$ is the spectral gap that measures the connectivity of the communication topology. These results then deliver an $\mathcal{O}{(N^{-(1+α)/2}+ m^{-(1+α)/2}+λ^{1+α} + ϕ_{\mathcal{S}})}$ in-average generalization bound, which is non-vacuous even when $λ$ is closed to $1$, in contrast to vacuous as suggested by existing literature on the projected version of D-SGD. Our theory indicates that the generalizability of D-SGD is positively correlated with the spectral gap, and can explain why consensus control in initial training phase can ensure better generalization. Experiments of VGG-11 and ResNet-18 on CIFAR-10, CIFAR-100 and Tiny-ImageNet justify our theory. To our best knowledge, this is the first work on the topology-aware generalization of vanilla D-SGD. Code is available at https://github.com/Raiden-Zhu/Generalization-of-DSGD.
format Preprint
id arxiv_https___arxiv_org_abs_2206_12680
institution arXiv
publishDate 2022
record_format arxiv
spellingShingle Topology-aware Generalization of Decentralized SGD
Zhu, Tongtian
He, Fengxiang
Zhang, Lan
Niu, Zhengyang
Song, Mingli
Tao, Dacheng
Machine Learning
This paper studies the algorithmic stability and generalizability of decentralized stochastic gradient descent (D-SGD). We prove that the consensus model learned by D-SGD is $\mathcal{O}{(N^{-1}+m^{-1} +λ^2)}$-stable in expectation in the non-convex non-smooth setting, where $N$ is the total sample size, $m$ is the worker number, and $1+λ$ is the spectral gap that measures the connectivity of the communication topology. These results then deliver an $\mathcal{O}{(N^{-(1+α)/2}+ m^{-(1+α)/2}+λ^{1+α} + ϕ_{\mathcal{S}})}$ in-average generalization bound, which is non-vacuous even when $λ$ is closed to $1$, in contrast to vacuous as suggested by existing literature on the projected version of D-SGD. Our theory indicates that the generalizability of D-SGD is positively correlated with the spectral gap, and can explain why consensus control in initial training phase can ensure better generalization. Experiments of VGG-11 and ResNet-18 on CIFAR-10, CIFAR-100 and Tiny-ImageNet justify our theory. To our best knowledge, this is the first work on the topology-aware generalization of vanilla D-SGD. Code is available at https://github.com/Raiden-Zhu/Generalization-of-DSGD.
title Topology-aware Generalization of Decentralized SGD
topic Machine Learning
url https://arxiv.org/abs/2206.12680