Subgradient sampling for nonsmooth nonconvex minimization

Jérôme Bolte (1) , Tam Le (1) , Edouard Pauwels (2, 3)

1 TSE-R - Toulouse School of Economics
2 IRIT-ADRIA - Argumentation, Décision, Raisonnement, Incertitude et Apprentissage
3 IRIT - Institut de recherche en informatique de Toulouse

Risk minimization for nonsmooth nonconvex problems naturally leads to firstorder sampling or, by an abuse of terminology, to stochastic subgradient descent. We establish the convergence of this method in the path-differentiable case, and describe more precise results under additional geometric assumptions. We recover and improve results from Ermoliev-Norkin by using a different approach: conservative calculus and the ODE method. In the definable case, we show that first-order subgradient sampling avoids artificial critical point with probability one and applies moreover to a large range of risk minimization problems in deep learning, based on the backpropagation oracle. As byproducts of our approach, we obtain several results on integration of independent interest, such as an interchange result for conservative derivatives and integrals, or the definability of set-valued parameterized integrals.

Keywords: Subgradient sampling; Stochastic gradient; Online deep learning; Conservative gradient; Path-differentiability

Domains: Optimization and Control [math.OC]; Machine Learning [stat.ML]

Jérôme Bolte, Tam Le, Edouard Pauwels. Subgradient sampling for nonsmooth nonconvex minimization. 2022. ⟨hal-03579383v1⟩