Part II - Algorithms
a quick overview under the optimization perspective, with a focus on linear relaxation based methods.

Click here: Slides for Part II - Algorithms (PDF)

tl;dr: We first focus on efficient incomplete verifiers that rely on relaxing the nonconvex activation functions in neural networks into convex domains, such as CROWN and DeepPoly. Then, we briefly discuss how to further tighten the bounds from incomplete verifiers through optimization (α-CROWN). Then, we focus on complete verifiers with a specific focus on the recent research trend of branch-and-bound (BaB) based verifiers such as (β-CROWN) relying on rapid and GPU-accelerated incomplete verifiers for each bounding step.


A neural network verification tutorial presented by Huan Zhang, Kaidi Xu, Shiqi Wang and Cho-Jui Hsieh

☕ Pudhina theme by Knhash 🛠️