Neurocomputing with High Dimensional Parameters - High Dimensional Neurocomputing: Growth, Appraisal and Applications - page 22

learning algorithm for its training are most popular in neural networks community.
The primary aim of the neural network is to learn input to output mapping, and the
learning algorithm achieves it by adjusting the parameters of the network, which are
weights and threshold values. As these weights and thresholds are real values in the
conventional neural network, it is also called real valued neural network (RVNN) or
conventional ANN or refers to neurocomputing with single dimension parameters.
A conventional ANN is a model that apes the real neuron described by many
researchers in the model description suggested time to time. The artificial neurons are
shown connected with links going from one layer to the one immediately succeeding
it, and some applications of neural networks, however, have had synapses that link the
neurons of the present layer with the ones not only of the immediately succeeding
layer, but also to the neurons that lie further up in the line (Lang and Witbrock
1988). The strength of the synapses (connections) is the synaptic strength, is the
weight associated with the connection. In the human brain, the weight is actually the
potential that controls the flow of electric impulses through the link. Each neuron has a
well defined aggregation function to process the integration of impinging signals and
an activation function that limits the output in predefined range. A typical activation
function that closely resembles the activation of real biological neuron is sigmoid
function as shown in Fig. 2.1 . A steepness factor was introduced to adjust the shape
of the activation function and tailor it to a form that closely resembles the actual
characteristic. A number of neurobiological studies and biophysics of computation
have inspired researchers to precisely state different aggregation function in literature.
Chapter 4 of this topic presents three new higher order neuron models based on these
studies.
ANN in real domain have limitations such as slow convergence and degree of accu-
racy achieved is normally lower, specially for many applications, which deal with
high dimensional signals. The easiest solution would be to consider a conventional
real domain neural network, where high dimensional signals are replaced by inde-
Fig. 2.1 Comparing the actual response curve of biological neuron with the mathematical function.
a The actual characteristic at the output of the neuron in the brain. b The characteristic described
by sigmoid function