Implementations of Neural Network Models in Silicon

The tasks of speech recognition, stereo vision, optical character recognition, video compression by vector quantization and many others are characterized by a desire to find the best match to an input data string in a large library of previously stored reference strings. In real applications the list may be of the order of a million words of 100 bits or more and some applications also require the best M matches (M around 5) to be passed to the next stage of data processing.