[FPBench] May 2nd Community Meeting: Zhongkui Ma on ReLU hull approximation
Ian Briggs
Hi everyone,
Next Thursday, May 2nd at 9:00-10:00 AM pacific time, will be the next FPBench Community Meeting.
If you are searching for the Zoom link, look no further: https://washington.zoom.us/j/92831331326
We are glad to welcome Zhongkui Ma from The University of Queensland to present on ReLU hull approximation.
Neural networks have offered distinct advantages over traditional techniques. However, the opaque neural networks render their performance vulnerable, in the presence of adversarial samples. This underscores the imperative of formal verification of neural networks, especially in the safety-critical domain. Deep neural networks are composed of multiple hidden layers with neurons, where mathematical operations stack linear and nonlinear (activation functions) operations. Using linear inequalities plays a crucial role in constraining and deducing the ranges of results from nonlinear operations. Considering correlations among input variables of multiple neurons leads to constraints known as multi-neuron constraints, posing a non-trivial high-dimensional challenge in computing the convex hull of functions. This work is dedicated to designing methods for computing multi-neuron constraints for the ReLU function to serve the robustness verification of neural networks. We have introduced a novel approach WRALU (Wrapping ReLU) for computing the convex hull of a ReLU function (published in POPL’24). This method significantly reduces computation time and complexity, constructing fewer constraints to achieve more precise approximations while handling higher dimensions.
We look forward to seeing everyone!
-Ian