remarkably similar to those described in the normalizing and template matching theories. Given information on its sensory perspective and images stored in its memory, the MPP performs precomparison processing to orient incoming images for compatibility with stored images.
Another hypothesis of perception with similar assumptions is feature detection or feature extraction theory. According to this formulation a pattern may be characterized as a configuration of elements or features which can be broken down into constituent subcomponents and put back together again. Recognition is a comparison process between lists of stored features (which when combined, constitute a pattern) and features extracted from incoming stimuli (Klatsky, 1975). An early theoretical AI model of the feature detection hypothesis was Pandemonium (Selfridge, 1966). This system performs a hierarchical comparison of low-level through higher-order features until the incoming pattern is recognized. More recent scene analysis paradigms have grown from similar assumptions that the raw scene must be "segmented" into regions, or edges of regions, out of which desired objects may be constructed (Arden, 1980; Barrow, private communication, 1980). Scene-analysis models developed on the basis of higher- order features of greater complexity than those proposed by Selfridge have achieved moderate success in limited environments. The major problem is that the system can only deal with familiar or expected input data. All categories within which items are recognized, must be explicitly defined by the programmer in terms of their subcomponents. This eliminates the possibility of recognition processes in novel environments.
Reviewed together, template matching and feature detection reflect the processes modeled by most AI imaging and pattern recognition research. Hence, current AI systems are incapable of handling new category construction and other advanced perceptual tasks which might be required in future space missions. This limitation suggests that an alternative approach to the problem of automated pattern recognition may be needed.
Despite abundant research supporting the existence of feature detectors in humans (Hubel and Wiesel, 1966; Lettvin et al., 1959), other evidence suggests that feature and template theory do not provide a complete explanation of recognition. The above approaches are regarded today as unsophisticated in their conception of how events are mentally represented, and erroneous in ignoring the problem of how representations are achieved. Experiments conducted by Franks and Bransford (1971) indicate that the human mental representation used for feature comparison may be prototypical and holistic rather than literal and elemental. That is, what is actually stored in memory is the product of an active construction, developed over time. In this view the cognitive system extracts and stores the converging "essences" of items to which it is exposed, and this abstraction is then utilized in the recognition process. The empha- sis is on conceptual representational construction and conceptually driven (top-down) processing, rather than matching and data-driven (bottom-up) processing. The advantage of a prototype approach to perception is that minor distortions or transformations within a limited range will not interfere with the recognition process.
The prototype approach may be considered in terms of two different aspects - the abstract analogical nature of representation and category or concept construction. With respect to machine intelligence, perhaps the closest approximation to the notion of prototypical representation is illustrated by Minsky's "frame" concept. A frame in Minsky's formulation is a data structure for representing a stereotyped situation (Minsky, 1975) and corresponds in many ways to the psychological notion of schema (Bartlett, 1961). Though not really analogical in nature, the frame conception contributes to scene analysis by permitting the system to access data in a top-down fashion and to utilize generalized information without relying on simplistic features. The frames, however, must be described within the system by a programmer and are relatively static. There is no capability for frame reorganization as a result of experience.
Consider now the second aspect of the prototype approach, the construction of abstract categorical representations. Category construction may be viewed as a brand of concept formation. Experimental evidence suggests that the formation of new conceptual categories is the result of a hypothesis generation and testing process (Levine, 1975) in which recursive operations are evoked which infer hypotheses about how a number of particulars are related and then test those hypotheses against feedback information from the environment. Some additional evidence suggests that a number of these hypotheses may be tested simultaneously (Bruner et al., 1956). The result is considered an abstract analogical representation capturing an essence which subsumes all the particulars. Since the hypothesis theory of concept formation typically has been considered in the context of conscious processes, it may seem somewhat far afield of perceptual processing. However, since perception itself has been described as an unconscious inferential process, it may be the case that similar underlying logical operations are at work in the formation of higher-order concepts, prototypes, and in perceptual construction. The precise nature of acquisition, how an "elegant" hypothesis is formed, is not clearly specified in any of these theories. (See section 3.3.3.)
Only a minimal amount of work has been done on AI approaches to the formation of new conceptual structures. A classic attempt was Winston's concept formation program in which a machine was taught through example to acquire new concepts (e.g., the architectural concept of "arch"). Using informational feedback from the programmer as to whether a particular example illustrated the concept or not, and by assessing the essential similarities and differences among
