Fast trainable pattern classification by a modification of Kanerva´s SDM model

A universal classifier called the modified sparse distributed memory (MSDM) is presented. Because the classifier is independent of the data distribution, the accuracy of classification does not require making any statistical assumptions. The reference addresses in MSDM are regular and dense instead of random and sparse as in sparse distributed memory. The random-address array is absent in an MSDM. The MSDM classifies patterns of small integers (smaller than 32) instead of binary numbers. Pattern similarity and address selection may be measured by a Euclidean rather than a Hamming distance. A learning procedure is first presented for the MSDM. Then a scheme for classifying multispectral images with large number of bands is described. The test results show that the MSDM is able to classify multispectral images with very high accuracy. The speed of classification by MSDM is totally competitive with the minimum distance classification (MDC), which currently remains one of the fastest traditional methods. However, MDC is not very accurate.<>