Solutions to Parallel and Distributed Computing Problems: Lessons from Biological Sciences

8.7 TASK MAPPING WITH NEURAL NETWORKS

A neural network (NN) is a computational structure inspired by the study of biological neural processing. A NN can be viewed as a directed graph wherein the nodes are artificial neurons. Neurons are processing elements seen as units that are similar to the neurons in a human brain. Hence, they are referred to as artificial neurons, or just neurons for simplicity. Edges of the directed graph represents the synapses, i.e., connections between the neurons. Numerical values, called weights, are attached to synapses.

The model of a neuron assumes multiple input and a single output. The input signals enter the neurons via synapses in which they are multiplied by weights. The weighted signals are added in the neurons. The output is produced by applying some function f, called activation function, to the weighted sum. In general, the weights and the functions are different in different synapses and neurons. The weights of the synapses can be modified adaptively in a process called learning.

The neurons can be connected in an arbitrary way. Different neural networks can be obtained from the general model given earlier, by specifying the interconnection structure, the activation function, and the learning algorithm.

8.7.1 Computing Using Neural Networks

To solve a problem using NN, it must be cast into a NN model. Researchers have suggested many different NN models. The Hopfield network, however, is often used in modeling a NN for optimization problems ...