Skip to Main Content
Conventional classification algorithms have already reached a plateau at the trade-off imposed by the bias due to the structure of the classifier and the variance due to the limited size of the training set. The latter may be alleviated by exploiting known constraints, including class and style priors, language models, statistical correlations between spatially proximate patterns, statistical dependence due to isogeny (common source) of patterns, and even information-theoretic properties of the representations that have evolved for symbolic patterns intended for communication. Another development that may lead to new applications of pattern recognition is more effective human intervention. The interplay of human and machine abilities requires models that are both human and computer accessible.