Parameter estimation and learning/classification threshold optimization applied to maxentropic adaptive pattern recognition

摘要

This paper discusses learning in pattern recognition in terms of parameter estimation. Methods of Estimation Theory are used to show that reinforcement learning is implemented by sequential parameter estimation which alters both a priori and spontaneously learned templates feature by feature. Spontaneous learning is handled as a problem in decision theory and optimum thresholds for learning or classification are developed in terms of Bayesian thresholds and decision statistics based on the mutual information of observables and templates.