A Learning Algorithm for Neural Networks and General Dynamic Systems

Abstract

In the past 10 years or so, machine learning has made great progress, in terms of both theories and applications. Among various machine learning methods, deep learning is one of the most widely used methods. Significant advance has been made in speech recognition, object recognition and detection as well as in many other applications using deep learning. In this talk, we present a learning algorithm that is mathematically equivalent to the well-known back-propagation learning algorithm in neural networks but have the following advantages over the back-propagation algorithm. (1) It does not require a feedback network to back-propagate the errors. This makes its implementations, especially implementations on silicon, much simpler. (2) It is biologically plausible as all information needed for synapses to adapt is available in a biological neuron. Hence, artificial neural networks indeed mimic biological neural networks. (3) It can be applied to general dynamic systems, that is, many dynamic systems can adapt in a way similar to neural networks. This property allows us to introduce “intelligence” into a large class of engineering systems. In particular, we consider adaptation in control systems using the proposed learning algorithm, including adaptive state feedback, adaptive PID controller, and model reference adaptive control.