Hessian Estimation via Stein's Identity in Black-Box Problems

Abstract

When the available information is noisy zeroth-order (ZO) oracle, stochastic approximation methods are popular for estimating the root of the multivariate gradient equation. Inspired by the Stein's identity, this work establishes a novel Hessian approximation scheme. We compare it alongside with second-order simultaneous perturbation stochastic approximation (2SPSA). On the basis of the almost sure convergence and the same convergence rate, 2SPSA requires four ZO queries, while ours requires three ZO queries. Moreover, 2SPSA requires two statistically independent perturbations and two differencing stepsizes, while ours requires generating one perturbation vector only and tuning one differencing stepsize only. Besides, the weighting mechanism for the Hessian estimate is generalized and the smoothness restriction on the loss function is relaxed compared to 2SPSA. Finally, we present numerical support for the reduced per-iteration ZO query complexity.

arXiv comment17 pages, 2 figures, accepted by Mathematical and Scientific Machine Learning 2021, to appear in Proceedings of Machine Learning Research

References in corpus (5)

• Fast large-scale optimization by unifying stochastic gradient and quasi-Newton methods
• Efficient Full-Matrix Adaptive Regularization
• Accelerated Stochastic Gradient-free and Projection-free Methods
• Hessian Inverse Approximation as Covariance for Random Perturbation in Black-Box Problems
• Error Bounds and Applications for Stochastic Approximation with Non-Decaying Gain