Training Neural Networks with Random Noise and Global Reward

Dr. Anthony Maida

Center for Advanced Computer Studies

Institute of Cognitive Science

University of Louisiana at Lafayette

 

Abstract

Synaptic weight-perturbation algorithms for supervised learning offer improved biological plausibility over backpropagation because of their reduced circuitry requirements. However, existing weight perturbation algorithms use biologically implausible controlled noise sources to determine their weight updates. Biological noise sources are less controllable and more indirect than those found in existing perturbation algorithms. The present work establishes via simulation studies that, although less controllable and more indirect, they can still support learning in a traditional neural network task.