Verification of performance of a neural network estimator

Abstract

This paper presents an approach for verifying performance of a feedforward neural net trained as a static nonlinear estimator, with a view to its use on commercial aircraft. The problem is important in context of safety-critical applications that require certification, such as flight software in aircraft. The algorithm presented here extends the previously published verification method developed for nets that approximate look-up tables. Through a suitable transformation, the problem is converted into verifying an approximation to a look-up table over a hyper-rectangular domain. Then, the previously developed technique is used. It is based on traversing a uniform testing grid and evaluating the error at its every node. The process results in guaranteed upper bounds on the error between the neural net estimate and the true value of the estimated quantity. The method allows deterministic verification of nets trained off-line to perform safety-critical estimation tasks.