«Discrimination Nets as Psychological Models* LAWRENCE W. BARSALOU Emory University GORDON H. B O W E R Stanford University Simulations of human ...»
COGNITIVE SCIENCE 8, 1-26 (1984)
as Psychological Models*
LAWRENCE W. BARSALOU
GORDON H. B O W E R
Simulations of human cognitive processes often employ discrimination nets to
model the access of permanent memory. We consider two types of discrimina-
tion nets--EPAM and positive-property-only nets--and argue that they have insufficient psychological validity. Their deficiencies arise from negative prop- erties, insufficient sensitivity to the discriminativeness of properties, extreme sensitivity to missing or incorrect properties, inefficiency in representing mul- tipLe knowledge domains, and seriality. We argue that these deficiencies stem from o high degree of test contingency in utilizing property information during acquisition and memory search. Discrimination nets are compared to other models that have less or no test contingency (e.g., PANDEMONIUM) and that thereby avoid the problems of discriminotion nets. We propose that under- standing test contingency and discovering psychologically valid ways to im- plement it will be central to understanding and simulating memory indexing in human cognition.
In reviewing i n f o r m a t i o n p r o c e s s i n g m o d e l s in the Annual Review of Psy- chology, S i m o n (1979, p. 378) s t a t e d, " T h e safest c o n c l u s i o n at the p r e s e n t t i m e is t h a t h u m a n L T M can p r o b a b l y be r e p r e s e n t e d as a n o d e - l i n k m e m o r y with an E P A M like index, b u t t h a t v a r i o u s a l t e r n a t i v e s a r e still o p e n for the detailed structure and organization of that memory." Subsequently, Lang- ley a n d S i m o n (1981, p. 363) i d e n t i f i e d " r e c o g n i t i o n b y m e a n s o f d i s c r i m i - *Work on this paper was supported by a NSF Graduate Fellowship to the first author and by NIMH Grant 13950 to the second. We are grateful to Irving Biederman, Jeffrey Farrar, Janet Kolodner, Christopher Riesbeck, Brian Ross, Daniel Sewell, and Edward Smith for help- ful comments. The final product does not necessarily represent their positions on its content.
Parts of this paper were presented at the Annual Stanford-Berkeley Experimental Psychology Conference, Stanford University, 1980. Requests for reprints should be addressed to Lawrence W. Barsalou, Department of Psychology, Emory University, Atlanta, GA 30322.
2 BARSALOU AND BOWERnation nets" as one of several "possible invariants of human cognition."
Cognitive scientists have used EPAM-like discrimination nets for the last 25 years to model how humans access information in long term memory. After describing E P A M nets and some of their applications, we argue that they have a number of deficiencies as models of long term m e m o r y access. We next consider simple variants of E P A M discrimination nets and argue that they too lack psychological validity. We then show that simple parallel indexing mechanisms such as P A N D E M O N I U M avoid the problems of discrimination nets. Finally, we consider each model's use of test contingency in property acquisition and m e m o r y search, and we argue that fundamental differences in test contingency underly these models' differences in performance.
E P A M (Elementary Perceiver and Memorizer) was developed by Feigenb a u m and Simon to account for "elementary hur,~,an symbolic learning processes" (e.g., Feigenbaum, 1963, p. 297; Simon & Feigenbaum, 1964). The central component of E P A M is an extremely efficient m e c h a n i s m - - a discrimination n e t - - t h a t sorts stimulus patterns to their correct actions on the basis of stimulus properties. Consider the example of an E P A M net in Figure 1. On the left is a list of stimulus-action pairs, in which letters represent properties of stimulus patterns and numbers represent actions. These could be the actual materials for a paired-associates task in which subjects learn to recall the number that goes with each letter string. We will use these pairs, however, to represent any kind of stimulus-action unit of behavior (e.g., object-category pairs in pattern recognition, cue-target pairs in m e m o r y retrieval, condition-action pairs in skilled performance).
The discrimination net on the right of Figure 1 generates the correct action for each stimulus. The non-terminal nodes in the net are tests of stimulus properties. If the stimulus EF is presented, it is first tested for having an E. Since the outcome is positive, EF is then tested for G. Failure on this test sorts EF to the terminal node that represents the correct action, 49. For AB, tests for E and then C fail, leading to the correct action, 62. We should note that this E P A M net cannot discriminate between stimuli that contain different permutations of the same properties (e.g., EG and GE). E P A M nets can easily make such discriminations, however, by making tests position specific. Instead of testing a stimulus for an E and a G in any position, a net could test it for an E in the first position and a G in the second position. An example of such a net is shown later at the b o t t o m of Figure 5.
An attractive feature of discrimination nets is their ability to learn discriminations. Consider how the net in Figure 1 would learn the new pair, EH-50. When first presented with EH, a test succeeds at E but not at G, and
62 94 49 57 49 50
the net erroneously produces 49. To differentiate the two stimuli sorted to the same terminal node, the net grows a new test node based on a property of E H that does not yet exist on its sorting path in the net. In this case, H has not yet been used and is added to the b o t t o m of the net where 49 once was. The old and new actions (49 and 50) are attached to the negative and positive branches. The net is now capable of responding correctly to both EF and EH.
A crucial c o m p o n e n t of E P A M is a "noticing o r d e r, " prespecified by the p r o g r a m ' s designer, that can control (a) the order in which the properties of a stimulus are tested, a n d / o r (b) the order in which the properties of a stimulus are added to its sorting path. One possibility for stimuli that are letter strings is a left-to-right order: When the properties of a stimulus are tested, they are tested from left to right; when a property is to be added to the sorting path of a stimulus, the left-most property not currently on the path is chosen. As we shall see, there are m a n y possible noticing orders.
E P A M nets have played a central role in computer programs whose primary goal is to simulate h u m a n cognitive processes. E P A M nets have successfully modelled a variety of verbal learning phenomena (Feigenbaum, 1963, 1965; Hintzman, 1968; Simon & Feigenbaum, 1964). Simon and Gilmartin (1973) used an E P A M net to recognize chess patterns in a simulation of chess perception, and G o l d m a n used an E P A M net as part of a speech production simulation in Schank's (1975) conceptual dependency system.
Although the efficiency of E P A M nets makes them attractive to artificial intelligence projects, we have some reservations about their psychological validity. We now present the reasons that underly our lack of confidence. ' 'J. R. Anderson and G. H. Bower (1973) also provide a critique of EPAM, but it largely addresses shortcomings related to their specific interests.
4 BARSALOU AND BOWER
The first problem with EPAM nets is their heavy reliance on negative properties as seen in the net in Figure 2. The stimulus for action 67 is defined in the net as not having an A, not having a B, and not having a D. It is defined solely by negative properties, that is, by properties it does not have. In any EPAM net, half the properties are negative, and only one stimulus can be recognized on the basis of all positive properties. We find this counter-intuitive and psychologically implausible for several reasons.
Figure 2. An EPAM net that exhibits the problems of negative properties, insufficient sensitivity to the discriminativeness of properties, and extreme sensitivity to missing or incorrect properties.
First, people do not seem to have extensive knowledge of negative properties explicitly represented in memory (although they can compute them when necessary). When subjects list the properties of stimuli, they rarely, if ever, provide negative properties (see the norms of Ashcraft, 1978;
Hemenway, 1981; Rosch, Mervis, Gray, Johnson, & Boyes-Braem, 1975).
Instead, people appear to primarily represent negative properties that deny normative expectations (e.g., the property of no vision for the concept of blindness).
Second, it seems counter-intuitive that people typically recognize things using negative information. This would be similar to walking into a room and identifying something as a chair because it does not fly and does
DISCRIMINATION NETS 5not have gills. People primarily appear to recognize things using positive information. Notably, the major models of stimulus structure proposed by psychologists do not employ negative properties (e.g., the geometric model of Shepard, 1962a, 1962b; the feature model of Tversky, 1977). Moreover, models like these do an acceptable job of accounting for pattern recognition data (e.g., Appleman & Mayzner, 1982; Gibson, 1965; Keren & Baggen, 1981; Krumhansl, 1982; McClelland & Rumelhart, 1981; Rumelhart & McClelland, 1982).
A third problem with negative properties is that nets using them do not credit people with as much knowledge of stimulus patterns as they undoubtedly have. Consider the stimulus for action 67 in Figure 2. The net could tell us nothing positive about the properties of that object; yet it would be surprising if someone who had correctly learned its action could not recognize any of its positive properties. The use of negative properties is so computationally powerful that not all positive properties in a knowledge domain need to be represented. As the line labelled positive properties in Figure 2 shows, this net represents only 7 of the 19 positive properties in the domain. Yet we would be quite surprised if someone who had learned all the pairs in Figure 2 could recognize less than half the properties. We suspect most of the properties would have become established in memory and could be recognized as having occurred within the domain.
Attempts have been made to some extent to remedy this lack-of-knowledge problem. The nets in Feigenbaum (1965) initially sorted stimuli to stimulus images instead of to actions. These images were presumed to build up with repeated experience until they eventually contained all a pattern's positive properties. To produce an action, a stimulus image was subsequently presented to the net and sorted to its action. But this appears to be quite redundant. Why should pattern recognition include both the classification of a percept and the subsequent classification of its memory representation, which is in some sense a copy? Representing a stimulus three times--once in the net that classifies the percept, once as a stimulus image, and once in the net that classifies the stimulus image--seems unnatural and inefficient. It would be more reasonable to use a single representation for identifying a pattern and for storing knowledge about its properties.
Hintzman (1968) proposed a second solution to the lack-of-knowledge problem. In one version of his SAL model, overlearning was permitted such that all a pattern's positive properties could eventually be encoded as tests in its net. But this addition violates the spirit of EPAM nets. It undercuts their claim to fast, efficient knowledge indexing based on very little information.
Clearly, if all the positive properties of a domain are eventually represented, there is no reason in the first place to represent negative properties. It would seem preferable to use some other, more efficient representation that contains only positive properties. Notably, human subjects, on occasion, do
6 BARSALOU AND BOWERlearn more information than is necessary to discriminate stimuli (Trabasso & Bower, 1968). Consequently, some form of overlearning is necessary to account for the acquisition of redundant discriminative properties.
A fourth problem with negative properties is that nets with such properties fail to capture the distinction between generating an action to a previously perceived stimulus and generating the same action to a completely novel stimulus. The net in Figure 2, for example, would produce the same action to unfamiliar stimuli such as WXY, GJK, and FEG as to the familiar stimulus QRS, and would have no way of determining which is the familiar pattern. Yet people would probably be much better at recognizing familiar patterns than unfamiliar ones that lead to the same action.
A fifth problem with negative properties is that nets with these properties form unintuitive equivalence classes. The patterns that could be sorted to 67 in Figure 2, for example, surely form a peculiar category. This category could include familiar patterns (e.g., QRS), new patterns whose properties have occurred in the domain before (e.g., FEG), and new patterns whose properties have never occurred in the domain (e.g., WXY). Quite often these categories violate the fundamental classification principle of maximizing within-category similarity and minimizing between-category similarity. As an illustration, consider how the net in Figure 2 would categorize the new stimulus GJK. According to the classification principle just stated, GJK should be classified with AJK, but the net instead classifies it with QRS. Thus, the use of negative properties leads to predictions about transfer performance that can be unintuitive and inconsistent with psychological data.
Insufficient Sensitivity to the Discriminativeness of Properties