Data Representations in Learning

Abstract

This paper examines the effect of varying the coarse-ness (or fine-ness) in a data representation upon the learning or recognition accuracy achievable. This accuracy is quantified by the least probability of error in recognition or the Bayes error rate, for a finite-class pattern recognition problem. We examine variation in recognition accuracy as a function of resolution, by modeling the granularity variation of the representation as a refinement of the underlying probability structure of the data. Specifically, refining the data representation leads to improved bounds on the probability of error. Indeed, this confirms the intuitive notion that more information can lead to improved decision-making. This analysis may be extended to multiresolution methods where coarse-to-fine and fineto-coarse variations in representations are possible.

We also discuss a general method to examine the effects of image resolution on recognizer performance. Empirical results in a 840-class Japanese optical character recognition task are presented. Considerable improvements in performance are observed as resolution increases from 40 to 200 ppi. However, diminshed performance improvements are observed at resolutions higher than 200 ppi. These results are useful in the design of optical character recognizers. We suggest that our results may be relevant to human letter recognition studies, where such an objective evaluation of the task is required.