Neural Networks: Evolution, Topologies, Learning Algorithms and Applications

Introduction

A neural network is a powerful data-modeling tool that is able to capture and represent complex input-output relationships (Haykin, 1999; Kumar, 2004). The motivation for the development of neural network technology stemmed from the desire to develop an artificial system that could perform intelligent tasks similar to those performed by the human brain. These tasks include understanding, cognition, perception, control of human body parts etc. Above all, the brain being the pivotal organ of the human body is also entrusted with the synchronization of these actions to avoid functional disorders with the help of the central nervous system, which by itself is an information processing entity. This multifaceted functionality of the human brain is attributed to the inherent massively parallel computing structure gifted by a highly dense structure of interconnected computing units referred to as neurons (approximately more than 10 billion in number), which evoke appropriate responses to external world signals (Rojas, 1996). These neurons or neuronal cells, which form the human nervous system, imbibe a series of biochemical reactions for the purpose of reception, storage, processing and transmission of information incident from the real world.

The neural cell body or soma houses the nucleus and is connected to treelike networks of nerve fibers called dendrites (Rojas, 1996). A single long fiber referred to as the axon extends from the soma. This axon eventually branches into strands and substrands to connect other neurons through junctions referred to as synapses. Complex biochemical reactions enable the transmission of signals from one neuron to another via the intermediate synapses when some transmitter compounds are released from the sending end of the junction. These compounds either raise or lower the electrical potential inside the body of the connected receiving soma. When the electric potential reaches above a threshold value, the signal pulse is sent from the sending soma to the receiving soma via the axon. At this point, the receiving cell is activated or fired. Thus, it is evident that the functioning of the nervous system is mainly carried out by the neurons (Rojas, 1996).

Artificial neural networks are targeted to model the information processing capabilities of the nervous system. The initial steps in this direction were envisaged by the introduction of the simplified neuron by McCulloch and Pitts in 1943 (McCulloch 1943). An artificial neural network comprises several computing primitives interconnected together in different topologies for processing and propagation of information. In analogy with the human nervous system, these primitives are referred to as neurons or nodes. The synapses or junctions between these neurons are represented by connection strengths (or weights), which weight or modulate the incident input signals. The nonlinear processing power of these neurons is guided by a characteristic activation/transfer function. To be precise, the impinging weighted signals from the anterior neurons sum up to generate a neuronal impulse by means of a transformation in the posterior neuron. In this way incident input information gets processed and propagated from the preceding neurons to the following ones. Eventually, relevant features of the information are learnt by the individual neurons by means of an application specific learning algorithm, which entails the adjusting of the interconnection weights until the desired precision is attained (Haykin 1999; Kumar, 2004). This learning of information content by the neurons of an artificial neural network is referred to as the training of the neural network. So much so forth, this trained information is manifested in the adjusted weights in a neural network.