Contractive neural networks are robust against noise
Robust Classification using Contractive Hamiltonian Neural ODEs
Our paper "Robust Classification using Contractive Hamiltonian Neural ODEs" is accepted for L-CSS. The early access is available here. In this work, we propose a neural ODE (a class of neural networks) that is contractive by design and we showed that it is robust against input small perturbations(e.g. white noise). Besides, we also guarantee non-exploding gradients that ensure numerical stability during the training.
Funding
This research is supported by the Swiss National Science Foundation under the NCCR Automation (grant agreement 51NF40 180545).