Linearly Separable Codes for Adaptive Threshold Networks

The significance of the coding assignment in linear threshold networks used for pattern recognition has been discussed. A heuristic procedure for generating linearly separable codes for a sample set of patterns has been outlined. Experimental results showing the recognition performance of codes so chosen have been shown. The performance of these codes is comparable to the 1-out-of-n code commonly used, but more optimum codes may exist. More work on the coding assignment problem is in order and should allow more efficient linear threshold networks in the pattern recognition application.