AI and Pattern Recognition

AI AND PATTERN RECOGNITION: PANEL DISCUSSION

A. Rosenfeld (Chairman), University of Maryland J. Feldman, University of Rochester L. Kana!, University of Maryland P. Winston, Massachusetts Institute of Technology

This panel deals with the relation­ ships between artificial intelligence and pattern recognition. It addresses such issues as: How do the fields differ? How should they relate to one another? In particular, does AI have anything to learn from PR?

Pattern recognition, classically, in­ volves the extraction of features and their subsequent analysis in feature space using statistical tools or clustering heuristics. There is certainly an inter­ esting class of problems that can be approached in this way, but for the prob­ lems of central interest in A.I., the feature space idea seems inappropriate, leaving feature extraction as an area of useful overlap between the fields.

An alternative pattern recognition paradigm is that of syntactic analysis, in which p~tterns are hierarchically decom­ posed into subparts, and recognition of an unknown pattern is basically a parsing process, where the pattern is first seg­ mented into primitive parts, and combin­ ations of these parts are matched against the rules of a "pattern grammar". This approach is somewhat more powerful, since it deals with structural descriptions rather than with unstructured feature vectors. However, from an AI standpoint, syntactic analysis is of limited useful­ ness if it makes no provision for using semantic knowledge about the patterns being analyzed.

On the other hand, there is something very appealing in the classical ideas of pattern recognition. The search for gen­ erality has become unpopular lately in AI. This is not because generality is undesir­ able, but because it is so incredibly difficult to achieve. The ideas of pattern recognition provide uniform al­ gorithms of great generality which require only changes in numerical values for their application to a new problem. It is worth considering how the paradigms of pattern recognition are relevant to achieving gen­ erality in areas such as image analysis.

It should also be pointed out that even in AI systems for tasks such as scene analysis, some use of pattern recognition methodologies is almost unavoidable. Such systems need to segment images into parts, just as in syntactic pattern recognition; to measure geometrical and textural pro­ perties of the parts, and to identify or classify the parts on the basis of these properties,just as in statistical pattern recognition. Advances in segmentation, feature extraction, and classification techniques are being made by both schools; they have more in common than is sometimes realized.

AI is centrally concerned with theoretical issues, but it has used a variety of practical tasks, e.g., in image and speech understanding, as test beds for studying machine perception. Pattern recognition has also dealt exten­ sively with these same tasks, but tradi­ tionally, has studied different classes of inputs, taken from areas such as docu­ ment reading, remote sensing, and medicine, rather than industrial automa­ tion. There are significant differences among these classes of scenes, but there are also many common problems. The over­ lap in applications is growing, and should lead to greater communication be­

tween the fields.