Lecture
This is a large class of systems whose architecture is analogous to the construction of nervous tissue from neurons. In one of the most common architectures — the multilayer perceptron with error backpropagation — the operation of neurons is simulated within a hierarchical network, where each higher-level neuron is connected via its inputs to the outputs of neurons in the layer below. The neurons of the lowest layer receive the values of the input parameters on the basis of which decisions must be made, the development of the situation must be forecast, etc. These values are treated as signals transmitted to the next layer, being weakened or amplified depending on the numerical values (weights) assigned to the inter-neuron connections. As a result, at the output of the topmost neuron a certain value is produced, which is regarded as the answer — the response of the entire network to the input parameter values entered.
In order for the network to be usable, it must first be trained on previously obtained data for which both the values of the input parameters and the correct answers to them are known. Training consists in selecting the weights of the inter-neuron connections that ensure the network's answers are as close as possible to the known correct answers.
The main drawback of the neural-network paradigm is the need to have a very large volume of training sample with correct answers, although modern knowledge repositories relatively easily make this possible. Another significant drawback is that even a trained neural network is a classic black box, which receives initial conditions at the input, processes them somehow, and produces a forecast at the output. Indeed, knowledge recorded as the weights of several hundred inter-neuron connections is practically impossible for a human to analyze and interpret.
Examples of neural-network systems in use — BrainMaker, NeuroShell, OWL.
Comments