Computational Tradeoffs of Optimization-Based Bound Tightening in ReLU Networks
This work addresses computational efficiency issues for researchers and practitioners using MILP to analyze or optimize ReLU networks, but it is incremental as it builds on existing MILP modeling approaches.
The paper tackles the tradeoff between bound tightness and computational effort in MILP models for ReLU neural networks, providing implementation guidelines based on network structure, regularization, and rounding.
The use of Mixed-Integer Linear Programming (MILP) models to represent neural networks with Rectified Linear Unit (ReLU) activations has become increasingly widespread in the last decade. This has enabled the use of MILP technology to test-or stress-their behavior, to adversarially improve their training, and to embed them in optimization models leveraging their predictive power. Many of these MILP models rely on activation bounds. That is, bounds on the input values of each neuron. In this work, we explore the tradeoff between the tightness of these bounds and the computational effort of solving the resulting MILP models. We provide guidelines for implementing these models based on the impact of network structure, regularization, and rounding.