The unsupervised learning is autonomous, without a priori. The traditional algorithm, proposed by Kohonen, is based on a strategy called "That the best gains". It supposes that the classes are linearly separable to the others.

In a layer of neurons of Kohonen, the number of neurons used is at least equal to the number of classes to be distinguished. Only the neuron which gives the best result for a given sample is adjusted.

The principle is based on that a neuron gives a degree of similarity between its vector weight and the vector of input of the sample. Thus, only the neuron which received the best initialization in order to recognize a class, will be improved.

After having presented all the samples, each neuron will have a vector weight which will represent the average vector of each class.

The components of the vector weight W of the "winner" neuron are adjusted by components of the input vector as follows:

then normalized by

The choice and the number of the initial vectors weight is useful such as the order of presentation of the samples. If some neurons have vectors weight too far from the classes presented, they will never be chosen, and thus will never be adjusted.