Mathematical Theory of Neural Networks
Eduardo D. Sontag, Héctor J. Sussmann · 1997
The work pursued under this grant dealt with artificial neural networks and other discrete/continuous models. New bounds were obtained for sample complexity for identification of static and dynamic concept classes defined by static and recurrent networks. Structural and system-theoretic properties were characterized, leading to effective tests for identifiability and other properties. Related models of hybrid systems were also studied; an equivalence problem for PL systems was shown to be decidable in polynomial time, and a general Maximum Principle was established for hybrid systems.