Rajesh Advani (B.E - Comps - TSEC)

An artificial neural network is defined as a structure composed of a number of interconnected units or artificial Neurons. Each unit has I/O characteristic and implements a local computation or Function. The output of any unit is determined by its I/O, its interconnection to other units, and possibly external inputs.

Any taxonomy to describe ANNs must begin with the identification of the following features:

NETWORK TOPOLOGIES:

In viewing network topologies and structures quantitatively as functions of unit interconnections, we can distinguish several concepts-

1. Recurrent Network (Feedback path exists)

2.Non-Recurrent Network (no closed interconnection path)

3. Layered, Hierarchical and another similarly

Structured Networks

4. Competitive interconnected structures.

Types 1 & 2 are mutually exclusive and types 3 & 4 may apply to either recurrent or non-recurrent networks.

ANN FUNCTIONS:

The desired behavior of the network provides another approach to distinguishing networks. For example, the desired function may be specified by enumerating a set of stable network states, or by identifying a desired network output as a function of the network inputs and current states. Popular examples of classifying ANNs by processing objective are:

1. The pattern associator (PA): This is implemented using feed-forward networks and it functionally relates patterns.

2. The content-addressable memory or associative memory model (CAM/AM): This is best exemplified by Hopfield model.

3. Self-Organizing Networks: These networks exemplify neural implementations of unsupervised learning in the sense that they typically self-organize input patterns into classes or clusters based on some form of similarity.

The mentioned models are explained below:

1] MULTILAYER FEEDFORWARD NETWORK STRUCTURE:

It is composed of a hierarchy of processing units, organized in a series of two or more mutually exclusive sets of neurons or layers. The first layer (input layer) serves as a holding site for the inputs applied to the network. The last layer (output layer) is the point at which the overall mapping of the network input is available.

Between these two extremes lie zero or more layer called hidden units; it is in the internal layers that additional remapping or computing takes place. Links or weights connect each unit in one layer only to those in the next higher layer. There is implied directionality in these connections. Each layer, based on its input, computes an output and propagates this information to the succeeding layer. Thus, from an architectural viewpoint, the feed-forward network allows parallelism within each layer but the flow of interlayer information is necessarily serial.

2] ASSOCIATIVE MEMORY MODEL: (Using Hopfield Network)

A Hopfield Network is a nonlinear network, put with the set of stored patterns.

3] COMPETITIVE AND SELF-ORANIZING NETWORKS:

Self-organizing approaches attempt to develop a network structure on the basis of the given data. One popular approach is 'clustering' or 'mode separation'. The objective is to design a mechanism that clusters data, perhaps by computing similarity. The cluster discovery capacity of such networks leads to the descriptor 'Self-Organizing'. In some applications, data fall into easily observed groups. The difficult case would be where the number of clusters and the separation between clusters are not visually obvious.

Clustering may be achieved through a number of strategies as follows;

1. Iterative approaches.

2. Hierarchical approach, which can be divided further as 'Merging of clusters' and 'Splitting of clusters'.