An Introduction to Formal Verification Methods for Neural Networks

This tech talk introduces the concept of formal verification for neural networks using intuitive examples and MATLAB^® tools. It explains two main verification approaches: the polyhedra method, which uses linear inequalities to define boundaries for the feasible input and output spaces but can be computationally expensive, and the interval method, which is computationally simpler but more conservative due to overestimation of bounds.

Modern algorithms such as DeepPoly and CROWN improve interval methods by propagating constraints backward through the network, tightening bounds, and reducing conservatism. Deep Learning Toolbox™ supports these techniques, enabling practical verification of complex networks to ensure robust and safe operation in critical systems.