Projectional Learning Laws for Differential Neural Networks Based on Double-Averaged Sub-Gradient Descent Technique

Isaac Chairez, Alexander Poznyak, Alexander Nazin, Tatyana Poznyak

Resultado de la investigación: Capítulo del libro/informe/acta de congresoContribución a la conferencia

Resumen

© 2019, Springer Nature Switzerland AG. A new method to design learning laws for neural networks with continuous dynamics is proposed in this study. The learning method is based on the so-called double-averaged descendant technique (DASGDT), which is a variant of the gradient-descendant method. The learning law implements a double averaged algorithm which filters the effect of uncertainties of the states, which are continuously measurable. The learning law overcomes the classical assumption on the strict convexity of the functional with respect to the weights. The photocatalytic ozonation process of a single contaminant is estimated using the learning law design proposed in this study.
Idioma originalInglés estadounidense
Título de la publicación alojadaProjectional Learning Laws for Differential Neural Networks Based on Double-Averaged Sub-Gradient Descent Technique
Páginas28-38
Número de páginas11
ISBN (versión digital)9783030227951
DOI
EstadoPublicada - 1 ene 2019
EventoLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) -
Duración: 1 ene 2019 → …

Serie de la publicación

NombreLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volumen11554 LNCS
ISSN (versión impresa)0302-9743

Conferencia

ConferenciaLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Período1/01/19 → …

Huella Profundice en los temas de investigación de 'Projectional Learning Laws for Differential Neural Networks Based on Double-Averaged Sub-Gradient Descent Technique'. En conjunto forman una huella única.

Citar esto