Revisiting Gradient Clipping: Stochastic bias and tight convergence guarantees

(2023) Revisiting Gradient Clipping: Stochastic bias and tight convergence guarantees.

Text 4069_revisiting_gradient_clipping_s.pdf - Accepted Version Download (961kB)

Abstract

Gradient clipping is a popular modification to standard (stochastic) gradient descent, at every iteration limiting the gradient norm to a certain value . It is widely used for example for stabilizing the training of deep learning models (Goodfellow et al., 2016), or for enforcing differential privacy (Abadi et al., 2016). Despite popularity and simplicity of the clipping mechanism, its convergence guarantees often require specific values of and strong noise assumptions. In this paper, we give convergence guarantees that show precise dependence on arbitrary clipping thresholds and show that our guarantees are tight with both deterministic and stochastic gradients. In particular, we show that (i) for deterministic gradient descent, the clipping threshold only affects the higher-order terms of convergence, (ii) in the stochastic setting convergence to the true optimum cannot be guaranteed under the standard noise assumption, even under arbitrary small step-sizes. We give matching upper and lower bounds for convergence of the gradient norm when running clipped SGD, and illustrate these results with experiments.

Item Type:	Conference or Workshop Item (A Paper) (Paper)
Divisions:	Sebastian Stich (SS)
Conference:	ICML International Conference on Machine Learning
Depositing User:	Sebastian Stich
Date Deposited:	30 Jun 2023 11:08
Last Modified:	30 Jun 2023 11:08
Primary Research Area:	NRA1: Trustworthy Information Processing
URI:	https://publications.cispa.saarland/id/eprint/3976

Actions

Actions (login required)

View Item