The neural network presented above is not of any interest because it has
not been trained (so it is not able to solve any particular problem).
This situation can be compared to a little baby, whose brain is fully developed
and ready for work but who is not able to do anything because it has not
experienced any stimulus. So a neural network without learning is analogous
to a human without education.
Therefore, a neural network must be trained to solve some particular
problem. The methodology of the training is analogous to the way you would
teach a child to read or to count, that is by presenting some number or
letter and by assiging the letters and numbers some values.
For example you could show a child of 5 years an image of an A and you could then tell him
that it's an "A'ye" with the best prononcation you could afford.
You will teach a neural network in exactly the same way, namely you will
feed our network with a set of numbers (in our library between 0 and 1), and
the network will give you a result in its output layer.
Since the weights of the connections in the network are initially in a random state, this result
will surely at the beginning not statisfy you, so you will change the weight
of some connections in order to obtain a better result.
You will change the weight of the connections in fact until you get the desired result (this is
the training stage).
Next you will feed the input layer of the network with other examples and continue adjusting weights, until eventually you obtain the desired output for each example.
The entire set of training examples must be shown to the network many times in order to get a satisfactory result. You would not expect a child to learn to read having seen each letter or word only once, similarly the network requires many examples.
After all of this training, your network is hopefully able to solve your problem, and we say that it has learned, and its `knowledge' is stored by all the different connection weights.
| Previous | Contents | Next |