Vowel pre-processing with a neurally based model

We assume that (1) nervous system activity is most usefully represented as the set of simultaneous individual neuron activities in a group of neurons, (2) different memory traces make use of the same synapses and (3) synapses associate two patterns of neural activity by incrementing synaptic connectivity proportional to the product of pre- and post-synaptic activity (a Hebbian rule). We extend this model by (1) introducting positive feedback of a set of neurons onto itself and (2) allowing the individual neurons to saturate. Positive feedback forces the pattern of neural activity into stable corners of a high dimensionality hypercube. The model has behavior reminiscent of 'categorical perception' in that large regions of initial neural activity will end in the same corner. We wish to demonstrate that this model can serve as an efficient pre-processer, which takes a noisy stimulus, a spoken vowel, and puts it in a noise free standard form. As a test, we used acoustic representations of nine spoken Dutch vowels. We apply the model, and show that after several thousand learning trials, input vowels, initially close together, are associated with separate stable corners.