Abstract
During the course of this project an electrical circuit was
developed. The resulting VLSI chip encodes an array of analog inputs
into a digital code. The chip performs analog-to-digital conversion,
intensity normalization and rough classification.
These properties could for example be useful for pre-processing in a speech
recognition system. The chip's inputs in such a system could be the
outputs of an analog bandpass filter array.
The output code is based on the principle of rank order coding,
capturing the order of firing across a neuron population.
A neuron array and a temporal winner-take-all structure are used to
produce the code. The implemented prototype has 31 analog inputs
and 128 bits digital output.
Theoretically, the chip would be able to classify 8.22 * 10^{33} (31!)
analog input patterns. Given that the input is a permutation of
a specific intensity profile used in the testing,
this information capacity is reduced to 3.7 * 10^{12} different
input patterns, due to noise.
This reduced information capacity gives reason to believe that the
speech phonemes could be classified by this chip.
However, this statement depends on an intensity profile
that might not be realistic for any phoneme.
This hypothesis has not been tried against a real speech
input within this project.