Next: G-LVQ Up: G-LVQa combination of Previous: Main objectives of

Kohonen's LVQ

As the neural algorithm to optimize, LVQ [1] is going to be used. LVQ is a supervised classification algorithm.

An LVQ network consists of an array of labeled vectors, which are conventionally called weight vectors. Each label corresponds to a class.

When training, each labeled vector in the training sample is submitted to the network. The closest weight vector (usually called the winning vector or neuron) is computed, and this vector is updated, in such a way that it gets closer to the input vector if both belong to the same class, and is moved away from it if they belong to different classes.

LVQ usually performs efficiently if the initial weight vectors are close to their final values. That is why some heuristic procedures are used for weight initialization, like Kohonen's SOM or any other vector quantization procedure, like k-means.

The number of weight vectors remains fixed during training, and is usually set prior to training to be equal to the number of classes, or to a fixed number of weight vectors per class. Several sets of initial weight vectors are tested, and the best solution is chosen.


jmerelo@
Mon Nov 21 11:09:09 MET 1994