This repository contains the code for running the experiments from the following paper:
Contracting Implicit Recurrent Neural Networks: Stable Models with Improved Trainability
We construct simple conditions that guarantee contraction in recurrent neural networks. The conditions take the form of a semidefinite constraint which we impose using the Burier-Monteiro approach to semidefinite programming. This results in a nonlinear equality constrained optimization problem that we solve using an augmented Lagrangian method.
We demonstrate the method on a simple simulated nonlinear example, and on a real gait prediction task.