Feedforward artificial neural networks constructed with the use of adaptive elements

E. N. Efimov, T. Ya. Shevgunov

Moscow Aviation Institute
(State University of Aerospace Technologies) 

Received August 13, 2012

Abstract. This article deals with feedforward artificial neural networks learned by supervised methods based on the error backpropagation. The artificial neural networks of mentioned type can be defined as the systems of interconnected adaptive elements transforming signals in two concurrent directions: either backward or forward. The key advantage of approach proposed is that the single adaptive element is no longer necessary to be a classical neuron or a layer of neurons, but it can be an arbitrary subsystem with any desirable transfer function. The presented method of neural network design is implemented in the developed software prototype along with the library of common adaptive elements. This paper also demonstrates the comparison of the results obtained by numerical simulation of the ultra-short-pulse radar response and by the classification of two random processes.
Keywords: neural network, backpropagation, adaptive element, signals and systems, gradient descent, Sage Math, Python.