Stability and Generalization of Learning Algorithms that Converge to Global Optima

Abstract

We establish novel generalization bounds for learning algorithms that converge to global minima. We do so by deriving black-box stability results that only depend on the convergence of a learning algorithm and the geometry around the minimizers of the loss function. The results are shown for nonconvex loss functions satisfying the Polyak-Łojasiewicz (PL) and the quadratic growth (QG) conditions. We further show that these conditions arise for some neural networks with linear activations. We use our black-box results to establish the stability of optimization algorithms such as stochastic gradient descent (SGD), gradient descent (GD), randomized coordinate descent (RCD), and the stochastic variance reduced gradient method (SVRG), in both the PL and the strongly convex setting. Our results match or improve state-of-the-art generalization bounds and can easily be extended to similar optimization algorithms. Finally, we show that although our results imply comparable stability for SGD and GD in the PL setting, there exist simple neural networks with multiple local minima where SGD is stable but GD is not.

arXiv comment27 pages, 5 figures

References in corpus (7)

• Understanding deep learning requires rethinking generalization
• On Large-Batch Training for Deep Learning: Generalization Gap and Sharp Minima
• Diverse Neural Network Learns True Target Functions
• Algorithmic stability and hypothesis complexity
• On the Generalization Properties of Differential Privacy
• Fast Rates for Empirical Risk Minimization of Strict Saddle Problems
• The Landscape of Deep Learning Algorithms

Cited by in corpus (3)

• Stochastic AUC Maximization with Deep Neural Networks
• Convergence of Adam for Non-convex Objectives: Relaxed Hyperparameters and Non-ergodic Case
• Distributed SGD Generalizes Well Under Asynchrony