A sequential decision approach to problems in pattern recognition and learning

This study is concerned with the problems in pattern recognition when input observations are taken sequentially with a proper stopping rule. The fact that sequential decision procedures (sequential probability ratio test and generalized sequential probability ratio test) can reduce the average risk in a monotonic fashion implies the gradual improvement of system performance during operation. The feature selection problem is also considered and both receptor (feature-extraction) and categorizer (decision) parts are treated as a whole. Before a terminal decision is made, a feedback path requests the categorizer to continue taking observations and properly adjust the stopping thresholds. At the same time, another feedback path feeds back the outcome of recognitions thus far obtained to the receptor for selecting a proper feature for the next observation. Using the proposed recognition scheme, the overall system performance has shown significant improvement during operation. Sequential recognition in random environment with and without an outside "supervision" is investigated, and the advantage of the proposed recognition system under "unsupervised" learning is discussed. Computer simulation of the recognition system with learning indicates very satisfactory results.