Reconciling Modern Deep Learning with Traditional Optimization Analyses: The Intrinsic Learning Rate

Author(s): Li, Zhiyuan; Lyu, Kaifeng; Arora, Sanjeev

To refer to this page use: http://arks.princeton.edu/ark:/88435/pr19p0d
DC FieldValueLanguage
dc.contributor.authorLi, Zhiyuan-
dc.contributor.authorLyu, Kaifeng-
dc.contributor.authorArora, Sanjeev-
dc.date.accessioned2021-10-08T19:50:45Z-
dc.date.available2021-10-08T19:50:45Z-
dc.date.issued2020en_US
dc.identifier.citationLi, Zhiyuan, Kaifeng Lyu, and Sanjeev Arora. "Reconciling modern deep learning with traditional optimization analyses: The intrinsic learning rate." Advances in Neural Information Processing Systems 33 (2020): pp. 14544-14555.en_US
dc.identifier.issn1049-5258-
dc.identifier.urihttps://proceedings.neurips.cc/paper/2020/file/a7453a5f026fb6831d68bdc9cb0edcae-Paper.pdf-
dc.identifier.urihttp://arks.princeton.edu/ark:/88435/pr19p0d-
dc.description.abstractRecent works (e.g., (Li \& Arora, 2020)) suggest that the use of popular normalization schemes (including Batch Normalization) in today's deep learning can move it far from a traditional optimization viewpoint, e.g., use of exponentially increasing learning rates. The current paper highlights other ways in which behavior of normalized nets departs from traditional viewpoints, and then initiates a formal framework for studying their mathematics via suitable adaptation of the conventional framework namely, modeling SGD-induced training trajectory via a suitable stochastic differential equation (SDE) with a noise term that captures gradient noise. This yields: (a) A new \textquotedblleft intrinsic learning rate\textquotedblright\ parameter that is the product of the normal learning rate η and weight decay factor λ . Analysis of the SDE shows how the effective speed of learning varies and equilibrates over time under the control of intrinsic LR. (b) A challenge---via theory and experiments---to popular belief that good generalization requires large learning rates at the start of training. (c) New experiments, backed by mathematical intuition, suggesting the number of steps to equilibrium (in function space) scales as the inverse of the intrinsic learning rate, as opposed to the exponential time convergence bound implied by SDE analysis. We name it the \emph{Fast Equilibrium Conjecture} and suggest it holds the key to why Batch Normalization is effective.en_US
dc.format.extent14544 - 14555en_US
dc.language.isoen_USen_US
dc.relation.ispartofAdvances in Neural Information Processing Systemsen_US
dc.rightsFinal published version. Article is made available in OAR by the publisher's permission or policy.en_US
dc.titleReconciling Modern Deep Learning with Traditional Optimization Analyses: The Intrinsic Learning Rateen_US
dc.typeConference Articleen_US
pu.type.symplectichttp://www.symplectic.co.uk/publications/atom-terms/1.0/conference-proceedingen_US

Files in This Item:
File Description SizeFormat