Reviews: A Unified View of Piecewise Linear Neural Network Verification

This paper presents a branch-and-bound based unified framework for piecewise linear neural network (PLNN) verification, and demonstrates how two existing methods (Reluplex and Planet) can be cast in these common terms. It explores other variations of various components of B-and-B, finding that a "smart branching (SB)" heuristic inspired by Kotler and Wong [12] results in a substantially faster method. My take on this paper is somewhat lukewarm, given the mix of strengths and weaknesses it has. STRENGTHS: The paper is very well-written, at least the first half of it. It motivates the problem well, sets the scope of the work (what it covers, what it doesn't), and summarizes the contributions in a meaningful way.