Strong mixed-integer programming formulations for trained neural networks
Joey Huchette
Postdoctoral researcher in the Operations Research group at Google Research
Abstract:
We present mixed-integer programming (MIP) formulations for high-dimensional piecewise linear functions that correspond to trained neural networks. These formulations can be used for a number of important tasks, such as: 1) verifying that an image classification network is robust to adversarial inputs, 2) designing DNA sequences that exhibit desirable therapeutic properties, 3) producing good candidate policies as a subroutine in deep reinforcement learning algorithms, and 4) solving decision problems with machine learning models embedded inside (i.e. the «predict, then optimize» paradigm). We provide formulations for networks with many of the most popular nonlinear operations (e.g. ReLU and max pooling) that are strictly stronger than other approaches from the literature. We corroborate this computationally on image classification verification tasks, where we show that our formulations are able to solve to optimality in orders of magnitude less time than existing methods.
Seminarios ISCI – Management Science & Analytics
Viernes 15 de marzo 2019
Desde las 13.30 a las 14.30 hrs
Sala de Asamblea, Piso 4 Beauchef 851
Almuerzos previa inscripción desde las 13:15 hrs
Consultas a seminarios@isci.cl