[Paper Review] The Anisotropic Noise in Stochastic Gradient Descent: Its Behavior of Escaping from Sharp Minima and Regularization Effects

Zhanxing Zhu, Jingfeng Wu+3 more|ePrints Soton (University of Southampton)|Mar 1, 2018

Stochastic Gradient Optimization Techniques94 citations

TL;DR

The paper analyzes how anisotropic noise in SGD, aligned with loss curvature, enhances escaping from sharp minima and yields regularization benefits, outperforming isotropic diffusion analogs.

ABSTRACT

Understanding the behavior of stochastic gradient descent (SGD) in the context of deep neural networks has raised lots of concerns recently. Along this line, we study a general form of gradient based optimization dynamics with unbiased noise, which unifies SGD and standard Langevin dynamics. Through investigating this general optimization dynamics, we analyze the behavior of SGD on escaping from minima and its regularization effects. A novel indicator is derived to characterize the efficiency of escaping from minima through measuring the alignment of noise covariance and the curvature of loss function. Based on this indicator, two conditions are established to show which type of noise structure is superior to isotropic noise in term of escaping efficiency. We further show that the anisotropic noise in SGD satisfies the two conditions, and thus helps to escape from sharp and poor minima effectively, towards more stable and flat minima that typically generalize well. We systematically design various experiments to verify the benefits of the anisotropic noise, compared with full gradient descent plus isotropic diffusion (i.e. Langevin dynamics).

Motivation & Objective

• Motivate understanding of SGD dynamics with unbiased noise and its impact on generalization.
• Introduce a general gradient-based optimization dynamics unifying SGD and Langevin dynamics.
• Derive a novel indicator measuring escaping efficiency via noise-curvature alignment.
• Establish conditions under which anisotropic noise outperforms isotropic noise for escaping minima.
• Empirically validate the benefits of anisotropic noise through multiple experiments on neural networks.

Proposed method

• Formulate a general gradient-based stochastic dynamics with unbiased noise that unifies SGD and Langevin dynamics.
• Derive an indicator based on the trace of the product of the Hessian and noise covariance Tr(H Sigma) to measure escaping efficiency.
• Analyze locally near minima using an Ornstein-Uhlenbeck approximation to relate escape behavior to noise structure.
• Prove propositions linking ill-conditioned Hessians and aligned anisotropic noise to superior escaping performance.
• Design and run experiments comparing SGD with various GLD variants (isotropic and anisotropic noise) on toy models and real datasets.

Experimental results

Research questions

• RQ1 How does the structure of SGD noise covariance Sigma, beyond its magnitude, affect escaping from minima?
• RQ2 Under what conditions does anisotropic noise aligned with the Hessian outperform isotropic noise in escaping sharp minima?
• RQ3 How is SGD noise covariance related to the loss landscape curvature near minima in neural networks?
• RQ4 Can anisotropic diffusion explain SGD’s regressive effect toward flat minima and improved generalization?

Key findings

• An indicator Tr(H Sigma) governs escaping efficiency, with higher values correlating with faster escape from minima.
• Anisotropic noise aligned with the Hessian can outperform isotropic noise in escaping sharp minima, especially for ill-conditioned Hessians.
• SGD noise covariance is related to the Hessian/Fisher information, implying alignment between gradient covariance and curvature near minima.
• In neural networks, SGD typically satisfies conditions that enable faster escape to flatter minima than isotropic diffusion methods.
• Experiments on toy models and real datasets (FashionMNIST, SVHN, CIFAR-10) show anisotropic SGD-like noise leads to flatter minima and better generalization compared to isotropic GLD variants.
• Isotropic noise does not significantly improve escape from sharp minima due to the anisotropic nature of loss landscapes in practice.