Abstract
This paper studies the asymptotic behavior of the constant step Stochastic Gradient Descent for the minimization of an unknown function, defined as the expectation of a non convex, non smooth, locally Lipschitz random function. As the gradient may not exist, it is replaced by a certain operator: a reasonable choice is to use an element of the Clarke subdifferential of the random function; another choice is the output of the celebrated backpropagation algorithm, which is popular amongst practioners, and whose properties have recently been studied by Bolte and Pauwels. Since the expectation of the chosen operator is not in general an element of the Clarke subdifferential of the mean function, it has been assumed in the literature that an oracle of the Clarke subdifferential of the mean function is available. As a first result, it is shown in this paper that such an oracle is not needed for almost all initialization points of the algorithm. Next, in the small step size regime, it is shown that the interpolated trajectory of the algorithm converges in probability (in the compact convergence sense) towards the set of solutions of a particular differential inclusion: the subgradient flow. Finally, viewing the iterates as a Markov chain whose transition kernel is indexed by the step size, it is shown that the invariant distribution of the kernel converge weakly to the set of invariant distribution of this differential inclusion as the step size tends to zero. These results show that when the step size is small, with large probability, the iterates eventually lie in a neighborhood of the critical points of the mean function.
Article PDF
Similar content being viewed by others
Avoid common mistakes on your manuscript.
References
Aliprantis, C.D., Border, K.C.: Infinite Dimensional Analysis: a Hitchhiker’s Guide. Springer, Berlin (2006). https://doi.org/10.1007/3-540-29587-9
Aubin, J.P., Cellina, A.: Differential inclusions, Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences], vol. 264. Springer, Berlin (1984). https://doi.org/10.1007/978-3-642-69512-4. Set-valued maps and viability theory
Aubin, J.P., Frankowska, H., Lasota, A.: Poincaré’s recurrence theorem for set-valued dynamical systems. Ann. Polon. Math. 54(1), 85–91 (1991). https://doi.org/10.4064/ap-54-1-85-91
Benaïm, M., Hofbauer, J., Sorin, S.: Stochastic approximations and differential inclusions. SIAM J. Control Optim. 44(1), 328–348 (2005). (electronic). https://doi.org/10.1137/S0363012904439301
Benveniste, A., Métivier, M., Priouret, P.: Adaptive algorithms and stochastic approximations, Applications of Mathematics (New York), vol. 22. Springer, Berlin (1990). https://doi.org/10.1007/978-3-642-75894-2. Translated from the French by Stephen S. Wilson
Bianchi, P., Hachem, W., Salim, A.: Constant step stochastic approximations involving differential inclusions: stability, long-run convergence and applications. Stochastics 91(2), 288–320 (2019). https://doi.org/10.1080/17442508.2018.1539086
Bolte, J., Daniilidis, A., Lewis, A., Shiota, M.: Clarke subgradients of stratifiable functions. SIAM J. Optim. 18(2), 556–572 (2007)
Bolte, J., Pauwels, E.: Conservative set valued fields, automatic differentiation, stochastic gradient method and deep learning. arXiv:1909.10300(2019)
Clarke, F.H., Ledyaev, Y.S., Stern, R.J., Wolenski, P.R.: Nonsmooth Analysis and Control Theory Graduate Texts in Mathematics, vol. 178. Springer, New York (1998)
Davis, D., Drusvyatskiy, D., Kakade, S., Lee, J.D.: Stochastic subgradient method converges on tame functions. Found Comput Math (20), 119–154. https://doi.org/10.1007/s10208-018-09409-5 (2020)
van den Dries, L., Miller, C.: Geometric categories and o-minimal structures. Duke. Math. J. 84(2), 497–540 (1996). https://doi.org/10.1215/S0012-7094-96-08416-1
Ermoliev, Y., Norkin, V.: Stochastic generalized gradient method for solving nonconvex nonsmooth stochastic optimization problems. Cybern. Syst. Anal. 34(2), 196–215 (1998). https://doi.org/10.1007/BF02742069. http://pure.iiasa.ac.at/id/eprint/5415/
Ermoliev, Y.M., Norkin, V.: Solution of nonconvex nonsmooth stochastic optimization problems. Cybern. Syst. Anal. 39(5), 701–715 (2003)
Faure, M., Roth, G.: Ergodic properties of weak asymptotic pseudotrajectories for set-valued dynamical systems. Stoch. Dyn. 13(1), 1250011,23 (2013). https://doi.org/10.1142/S0219493712500116
Folland, G.: Real Analysis: Modern Techniques and Their Applications. Pure and Applied Mathematics: A Wiley Series of Texts, Monographs and Tracts. Wiley, Hoboken (2013). https://books.google.fr/books?id=wI4fAwAAQBAJ
Has’minskiı̆, R.Z.: The averaging principle for parabolic and elliptic differential equations and Markov processes with small diffusion. Teor. Verojatnost. i Primenen. 8, 3–25 (1963)
Ioffe, A.D.: An invitation to tame optimization. SIAM J. Optim. 19(4), 1894–1917 (2009). https://doi.org/10.1137/080722059
Kakade, S., Lee, J.D.: Provably correct automatic sub-differentiation for qualified programs. In: Bengio, S., Wallach, H., Larochelle, H., Grauman, K., Cesa-Bianchi, N., Garnett, R. (eds.) Advances in Neural Information Processing Systems. http://papers.nips.cc/paper/7943-provably-correct-automatic-sub-differentiation-for-qualified-programs.pdf, vol. 31, pp 7125–7135. Curran Associates, Inc (2018)
Kushner, H.J., Yin, G.G.: Stochastic Approximation and Recursive Algorithms and Applications, Applications of Mathematics (New York), 2nd edn., vol. 35. Springer, New York (2003). Stochastic Modelling and Applied Probability
Lebourg, G.: Generic differentiability of Lipschitzian functions. Transactions of the American Mathematical Society 256, 125–144 (1979). http://www.jstor.org/stable/1998104
Majewski, S., Miasojedow, B., Moulines, E.: Analysis of nonsmooth stochastic approximation: the differential inclusion approach. arXiv:1805.01916(2018)
Meyn, S., Tweedie, R.L.: Markov Chains and Stochastic Stability, 2nd edn. Cambridge University Press, New York (2009)
Mikhalevich, V., Gupal, A., Norkin, V.: Methods of nonconvex optimization. Nauka (1987)
Norkin, V.: Generalized-differentiable functions. Cybern. Syst. Anal. 16, 10–12 (1980). https://doi.org/10.1007/BF01099354
Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic Differentiation in PyTorch. In: NIPS-W (2017)
Roth, G., Sandholm, W.H.: Stochastic approximations with constant step size and differential inclusions. SIAM J. Control Optim. 51(1), 525–555 (2013). https://doi.org/10.1137/110844192
Ruszczyński, A.: Convergence of a stochastic subgradient method with averaging for nonsmooth nonconvex constrained optimization. Optim. Lett. 14. https://doi.org/10.1007/s11590-020-01537-8 (2020)
Acknowledgements
The authors wish to thank Jérôme Bolte and Edouard Pauwels for their inspiring remarks. This work is partially supported by the Région Ile-de-France.
Funding
The work of the third author is supported by the Région Ile-de-France.
Author information
Authors and Affiliations
Corresponding author
Ethics declarations
Computing interest
Not applicable.
Additional information
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Availability of data and materials
Not applicable.
Rights and permissions
About this article
Cite this article
Bianchi, P., Hachem, W. & Schechtman, S. Convergence of Constant Step Stochastic Gradient Descent for Non-Smooth Non-Convex Functions. Set-Valued Var. Anal 30, 1117–1147 (2022). https://doi.org/10.1007/s11228-022-00638-z
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s11228-022-00638-z
Keywords
- Clarke subdifferential
- Backpropagation algorithm
- Differential inclusions
- Non convex and non smooth optimization
- Stochastic approximation