Counting linear regions for neural networks verification

-- Guillaume Charpiat & Julien Girard-Satabin (Université Paris-Saclay)

Time: 10:30 -- Location:

Neural networks are becoming an increasingly popular class of programs. However, their relative lack of safety presents an obstacle to their wide acceptance yet to be tackled. To approximate complex behaviours, neural networks rely on piece-wise linear activation functions in their control flow. This piece-wise linear behaviour yield regions on the input space where the neural network behaves linearly. We devised a method exploiting those linear regions to speed up formal verification on neural networks. This method rely on the counting of linear regions. We present our method and an algorithm to enumerate the linear regions of a neural network.

Category: seminars
Tags: Combi seminar combinatorics