Information Technology Reference
In-Depth Information
description of any neuron model, one needs to know how the input signals are aggre-
gated and processed to obtain the output.
ANN, a simplified model of the biological neural network, is a massively parallel
distributed processing systemwhich is made up of highly interconnected neural com-
puting units (neurons) having ability to learn and generalize. Generalization refers to
the neural network producing appropriate outputs for inputs not encountered during
training. The learning and generalization capabilities of neural networks make them
possible to solve the complex problems which are currently intractable. The impor-
tant process in artificial neuron is forming a unit net potential from impinging signals.
There are various ways to aggregate the input values to get the unit potential value
like additive, subtractive, multiplicative, polynomial, rational, etc. The summation
of inputs in neuron played a vital role in construction of an artificial neuron. Such
neuron model when used to solve the real-life problem may require a large number
of neurons in the network which means that the complexity of the network and its
computational burden will be extensively increased. This problem more worsens
when dealing with high- dimensional applications. This problem can be overcome
by the consideration of various factors, such as architecture, node's functionality,
learning rule, and training sets. Their appropriate choice is necessary for an efficient
design of the ANN. But, there are three leading subjects to trim down the complex-
ity and computational burden of network along with efficient learning and superior
results.
•
Reducingthenumberofneurons inANNbyintroducinghigher-orderneurons.
The higher-order neuron may produce better learning and generalization perfor-
mance with reasonably less number of connected nodes.
•
Fastenthe trainingprocessbyselectingtheefficient learningalgorithm.
The learning speed of ANN depends upon the nature of the learning algorithm.
Apart from straightforward, simple, and conventional backpropagation algorithm
there are many learning algorithm which provide much faster convergence.
•
Implementationofneuralnetworks inhigherdimensions.
The learning and architectural complexity of ANN also depends upon the com-
plexity of the application concerned. When dealing with high-dimensional data,
it is better to consider neurons which accepting high-dimensional data as single
cluster. This drastically reduces the number of connected nodes as compared in
the conventional ANN.
In the last decade, many researches in the field of ANN have been directed toward
the evolution of superior architecture for neural system for efficient learning and
better analysis of high-dimensional data. This chapter is mainly concerned with
mathematical modeling of higher-order neuron and their implementation in complex
domain, which is an important aspect in studying the fundamental principal of infor-
mation processing. Three example neuron models are presented here to demonstrate
the motivation of introduction of higher-order neuron modeling.
Search WWH ::
Custom Search