Robustly and Optimally Controlled Training Of neural Networks I (OCTON I)

This project aims at developing novel data training methods for network of function approximators (such as neural networks) based on robust and optimal control theory. The main idea is to utilize approximate neural tangent kernel parametrization in order to dynamically constrain training objectives. In this project we will develop novel methods that accounts for non-traditional training objectives (other than mean square prediction error) and corrupted data sequence. The latter claims for robustification. Conservativeness, stability of training, guaranteed rate of convergence, scalable numerical optimization routines will be developed.