Correspondence between neuroevolution and gradient descent.
Journal
Nature communications
ISSN: 2041-1723
Titre abrégé: Nat Commun
Pays: England
ID NLM: 101528555
Informations de publication
Date de publication:
02 11 2021
02 11 2021
Historique:
received:
26
04
2021
accepted:
04
10
2021
entrez:
3
11
2021
pubmed:
4
11
2021
medline:
24
12
2021
Statut:
epublish
Résumé
We show analytically that training a neural network by conditioned stochastic mutation or neuroevolution of its weights is equivalent, in the limit of small mutations, to gradient descent on the loss function in the presence of Gaussian white noise. Averaged over independent realizations of the learning process, neuroevolution is equivalent to gradient descent on the loss function. We use numerical simulation to show that this correspondence can be observed for finite mutations, for shallow and deep neural networks. Our results provide a connection between two families of neural-network training methods that are usually considered to be fundamentally different.
Identifiants
pubmed: 34728632
doi: 10.1038/s41467-021-26568-2
pii: 10.1038/s41467-021-26568-2
pmc: PMC8563972
doi:
Types de publication
Journal Article
Research Support, Non-U.S. Gov't
Research Support, U.S. Gov't, Non-P.H.S.
Langues
eng
Sous-ensembles de citation
IM
Pagination
6317Informations de copyright
© 2021. The Author(s).
Références
Biosystems. 1994;32(3):171-82
pubmed: 7919114
Eur Phys J E Soft Matter. 2018 May 14;41(5):59
pubmed: 29748868
Phys Rev E Stat Nonlin Soft Matter Phys. 2007 Jul;76(1 Pt 1):011507
pubmed: 17677450
J Chem Phys. 2007 Aug 28;127(8):085106
pubmed: 17764305
J Chem Phys. 2007 Oct 21;127(15):154101
pubmed: 17949126
J Chem Phys. 2010 May 21;132(19):194102
pubmed: 20499946
Phys Rev Lett. 1989 Jan 23;62(4):361-364
pubmed: 10040213
Evol Comput. 2002 Summer;10(2):99-127
pubmed: 12180173
Neural Netw. 2015 Jan;61:85-117
pubmed: 25462637
Phys Rev E. 2020 May;101(5-1):052604
pubmed: 32575260
Phys Rev Lett. 1987 Jan 12;58(2):86-88
pubmed: 10034599
J Chem Phys. 2013 Jun 28;138(24):244111
pubmed: 23822231
Phys Rev Lett. 2004 Jan 23;92(3):035504
pubmed: 14753886