Feature selection in codebook based methods provides high accuracy

Despite the higher efficiency obtained by some algorithms (quasi-Newton methods, cascade correlation, etc.) in feedforward neural networks, faster learning methods such as those based on codebook vectors are still needed. We propose to perform feature selection in codebook based methods to improve their accuracy. However, we define a neural network with an exact and fast parallel implementation of the nearest network rule which allows previous feature selection by means of a pruning method. Moreover, we apply this feature selection algorithm upon another codebook based classifier - the Kohonen´s linear vector quantization