1.3 KiB
First Pass
Category: This is a methods paper.
Context: This paper proposes a way of using mixed integer linear programming (MILP) to evaluate properties of neural networks.
Correctness: Formal
Contributions: They do nifty things with bounds tightening and presolving that makes their solver very fast compared to the state of the art or Reluplex. They also talk about stable and unstable neurons.
Clarity: They have a really good explanation of what a MILP problem is and how one might encode a neural network as one.
Second Pass
What is the main thrust? The main thrust is their new solving method of MILPs for neural networks. With their method, neural networks can have their neurons analyzed to prove whether or not the network is robust to input perturbations. This is especially important for classifiers, who need to know if there are sneaky nonlinearities that can be harmful to a built system (like a glitch). This method of bounds tightening and MILP usage makes their solver much faster and therein more capable to handle large networks.
What is the supporting evidence? They have a whole bunch of experimental results.
What are the key findings? MILPs and bound tightening is very good!
Third Pass
Recreation Notes:
Hidden Findings:
Weak Points? Strong Points?