Noninvertibility in neural networks
Ramiro Rico-Martı́nez, Ioannis G. Kevrekidis, R.A. Adomaitis · 2002
A method for assessing certain validity aspects of predictions made by neural networks used to approximate continuous (in time) dynamical systems is presented. This method searches for noninvertibility (nonuniqueness of the reverse time dynamics) of the fitted model, an indication of breakdown of proper dynamical behavior. It is useful for computing bounds on the valid range of network predictions.>