In this work we investigate the use of a binary CMM (Correlation Matrix Memory) neural network for pattern classification. It is known that a k-NN rule is applicable to a wide range of classification problems but it is slow, and that the CMM is simple and quick to train, and has highly flexible and fast search ability. We combine the two techniques to obtain a generic and fast classifier which uses a CMM for storing and matching a large amount of patterns efficiently, and the k-NN rule for classification. To meet requirements of the CMM, a robust encoder has been developed to convert numerical inputs into binary ones with the maximally achievable uniformity. Experimental results on several benchmarks show our method can be over 4 times faster than the simple k-NN method with less than 1percent lower classification accuracy.
展开▼
