Search for: All records

Two pretrained neural networks are deemed (approximately) equivalent if they yield similar outputs for the same inputs. Equivalence checking of neural networks is of great importance, due to its utility in replacing learning-enabled components with (approximately) equivalent ones, when there is need to fulfill additional requirements or to address security threats, as is the case when using knowledge distillation, adversarial training, etc. In this paper, we present a method to solve various strict and approximate equivalence checking problems for neural networks, by reducing them to SMT satisfiability checking problems. This work explores the utility and limitations of the neural network equivalence checking framework, and proposes avenues for future research and improvements toward more scalable and practically applicable solutions. We present experimental results, for diverse types of neural network models (classifiers and regression networks) and equivalence criteria, towards a general and application-independent equivalence checking approach.