Neural networks

We have seen a need for systems to learn (a process which involves feedback) - but how can this be implemented? Can we in fact produce systems that are 'intelligent'? We could use a computer, suitably programmed - which in effect has one (or a few) 'processing' elements. However, even modern computers are not that advanced - perhaps it would be better to develop systems more like the most powerful learning systems - brains.

A brain comprises simple processing elements - called neurons - which act rather slowly - typically doing 1000 operations a second, whereas a computer does many millions. However, the brain has billions of neurons - connected together in a network - the net result of which is much more powerful than a normal computer.

Thus we 'borrow' from nature and try to develop artificial neural networks ANNs - being many neurons connected together. There are many ways of implementing these, but one method is to have neurons which multiply each input by its 'weight' being a value associated with the connecting link to the neuron, and the neuron output is the sum of all such weights. The neuron output may well provide the input to other neurons.

So that such a network can generate the 'right' results for any system, the correct weights are needed, but finding them is non trivial. So a 'training set' of inputs and correct answers is provided.

For each set, the inputs are passed to the network, the outputs are calculated, and any error between this calculated outputs and the expected outputs are used to adjust the weights - it is a feedback process.
 

<< Learning | Virtual Reality and HCI >>