|
|
|
|||||||||||||||||||||||||||||||||||||||||||||||
|
|
|
|
|
|
||||||||||||
|
||||||||||||
Trends
in Cognitive Sciences
Volume
3, Issue 4, 1 April 1999, Pages 123-124
|
|
|
|
This Document |
|
|
|
|
|
|
|
|
Full Text + Links |
|
|
|
|
|
|
|
|
|
|
Actions |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
||
PII: S1364-6613(99)01292-9
Copyright © 1999 Elsevier Science Ltd. All
rights reserved.
Update
What can
neuropsychology tell us about category learning?
Barbara J. Knowlton
, 
Department of Psychology, UCLA, Los Angeles, CA 90095, USA.
Available online 18 June 1999.
|
Referred to by: |
Math modeling, neuropsychology, and category learning:;
Response to B. Knowlton (1999), Trends in Cognitive Sciences, Volume
3, Issue 4, 1 April 1999, |
Author Keywords: Categorization;
Category learning; Implicit learning; Mathematical models; Amnesia;
Neuropsychology
Subject-index terms:
Neuroscience
Article Outline
The learning of categories has been a popular and contested topic in cognitive science. One reason for this interest is its importance and widespread application. For example, we can readily classify an unfamiliar painting as being the work of a known artist based on our previous experience with that artist's work. Experts can perform classification tasks ranging from sexing one-day-old chicks to diagnosing illnesses to detecting sonar patterns of enemy submarines.
Another reason for the interest in the topic is that there is a clear dichotomy in theories of category learning. On one hand, category learning appears to be a perfect example of implicit learning. It is not necessary to try to learn about the characteristics of a category in order to be able to categorize items: previous exposure to category exemplars is sufficient to enable people accurately to classify new items into the category[1 and 2]. Furthermore, people are often extremely poor at describing the basis for their categorization ability. Even those of us who are not schooled in music theory could classify an unfamiliar piece of music as `sounding like Mozart', though we might not have the insight to say exactly what this means.
According to a contrasting view, categorization tasks are not accomplished by reliance on a body of implicitly acquired category-level knowledge, but rather by reliance on comparison to category exemplars themselves[3, 4, 5 and 6]. It is thus not surprising that individuals can classify new items well without being able to articulate category criteria because this view holds that they are not using this information anyway. Although there are many different types of exemplar-based models, they all suggest at root that people classify an item into a category based on its similarity to other items in that category. Thus, categorization relies on memory for exemplars, presumably the same memory traces that allow for the recognition of those exemplars. Although exemplar-based models might not seem intuitive (do we really compare a painting to our specific memories of all other paintings that we have seen in our lifetime in order to guess that it is a David Hockney?), they have been shown to do an excellent job of accounting for the findings of category-learning experiments. Furthermore, they have the advantage of parsimony, because one does not need to hypothesize a mechanism for the acquisition and storage of category-level information in addition to exemplar information.
Given that both of these alternatives are quite plausible when one considers data from normal subjects, a neuropsychological approach could help to differentiate between the two. In work done in collaboration with Larry Squire, we found that amnesic patients were able to learn to classify items as well as intact subjects were in a variety of paradigms, despite the amnesics' severely impaired declarative knowledge for the exemplars used to teach the category[7, 8 and 9]. Because of their damage to medial temporal and diencephalic regions subserving declarative memory storage, we reasoned that these patients should exhibit category-learning deficits if category learning depends on the same knowledge of exemplars that gives rise to recognition performance.
In a recent paper, Nosofsky and Zaki[10] suggest that the dissociation between classification and recognition shown by amnesic patients does not necessarily refute exemplar models of category learning. They demonstrated that an exemplar-based model can account for the results of Knowlton and Squire [8], in which amnesic patients exhibited intact learning in a dot-pattern classification task despite impaired recognition memory for dot patterns. This study used a procedure derived from that of Posner and Keele [1], in which subjects viewed exemplars that were highly distorted versions of a randomly generated prototype dot pattern. In the test, subjects classified new dot patterns that were either the category prototype, low distortions of this prototype, high distortions of this prototype, or distracters that were highly distorted versions of other randomly selected prototypes.
Using identical materials to Knowlton and Squire[8], Nosofsky and Zaki found that their exemplar-based model performed as well as normal subjects and amnesic patients on the classification test, even when exemplar knowledge in the model was degraded to the point that recognition memory was no better than that of the am-nesic patients [10]. Qualitatively, this result was obtained because the relationship between exemplar knowledge and categorization ability is not linear: in the model, a small amount of exemplar knowledge can go a long way to give rise to classification ability that is not significantly different from normal.
If both the multiple-system view and the exemplar-based view can account for the dissociation between recognition and categorization seen in amnesic patients, might neuropsychological data further inform the debate? One strong prediction made by the multiple-system view is that category learning could proceed normally if absolutely no exemplar knowledge is present. An exemplar-based view predicts that no category learning can take place if no exemplar knowledge is acquired. Testing a patient with no ability to acquire declarative memories could therefore shed light on this question.
One such patient is E.P., a man who became profoundly amnesic following a bout of herpes encephalitis that resulted in virtually complete bilateral damage to the medial temporal lobes[11 and 12]. Although most amnesic patients described in the literature have some residual declarative memory ability, E.P. performs essentially at chance on all standard tests of declarative memory. Not surprisingly, E.P. performed at chance on a recognition test of dot patterns in which matched control subjects performed at about 95% correct. However, E.P. performed as well as these control subjects did on dot pattern classification tests [13]. These results therefore strongly support the idea that category learning does not necessarily rely on declarative memory for exemplars. Nosofsky and Zaki [10] do make the point that recognition memory tests such as the one used by Squire and Knowlton [13] might not be sensitive enough to detect underlying exemplar knowledge, especially if performance is affected by strategic factors such as a subject `giving up' during the test. Nosofsky and Zaki suggest that E.P. might have possessed some unde-tected exemplar knowledge that allowed him to perform classification normally. However, this point loses sight of the profundity of E.P.'s amnesia: not only did he perform at chance when discriminating old and new dot patterns, he consistently had no memory of seeing any dot patterns before. Even more astounding, he failed to recognize the experimenter even after a few dozen visits [13]. Thus, it seems improbable that a more sensitive probe of his recognition memory could reveal some minute amount of specific exemplar knowledge, which could then give rise to his normal (in fact, numerically slightly superior) classification performance.
A second prediction made by the multiple-systems view is that declarative-memory performance and classification performance can be double- dissociated. Knowlton, Mangels and Squire[14] showed that amnesic patients were able to perform normally on a probabilistic classification task, in which a series of cues were each probabilistically associated with one of two outcomes (sunny or rainy weather). The amnesic patients were impaired, as expected, on a test of declarative memory for the testing episode. Patients with Parkinson's disease, which affects primarily the basal ganglia, exhibit the opposite pattern of results. They remember the training episode normally, as assessed by a recognition memory test, yet they are severely impaired at learning the cueÆoutcome associations in the classification task. Although these data are consistent with a multiple-systems view, Nosofsky and Zaki suggest that this double dissociation could have been obtained if classification performance in the Parkinsonian patients was depressed because of a sub-optimal response strategy in this group. Thus, they argue, although their exemplar knowledge was normal, their response rules gave rise to a deficit.
One prediction that could be made from a response-strategy explanation is that the patients with Parkinson's disease should exhibit a deficit throughout training compared with normal subjects. In the results obtained by Knowlton, Mangels and Squire[14] and in a previous study [9], normal subjects surpassed the performance of amnesic patients when both groups received extended training (more than 50 trials). One possible explanation for this finding is that enough declarative knowledge is eventually acquired to allow subjects to use this information. Supporting this idea is the fact that normal subjects do exhibit some flexible knowledge of the cueÆoutcome relationships after 50 trials of training [15]. If declarative knowledge of cueÆoutcome relationships plays a major role in performance after extended training, then one would predict that patients with Parkinson's disease could perform quite well if given enough training. If, on the other hand, Parkinsonian patients exhibit a general strategic deficit at the response selection stage, this deficit would persist throughout training.
In fact, the data suggest that patients with Parkinson's disease do catch up with normal performance with extended training. In the results of Knowlton et al.[14] there was no significant difference between controls and patients with Parkinson's disease on trials 100Æ150 of the probabilistic classification task, although there was a highly significant difference over the first 50 trials. Although it is impossible to rule out a strategic deficit based on this non-significant difference in late training, the available data do seem more consistent with a learning deficit early in training in the Parkinsonian patients.
Much of the appeal of exemplar-based models lies in their parsimony. If a single knowledge-base can elegantly account for both categorization and recognition, why hypothesize multiple knowledge-bases? However, looking at parsimony at the behavioral level alone can be misleading. I would argue that by taking a cognitive neuroscience perspective, and thus acknowledging that the brain is composed of multiple systems that exhibit plasticity, it seems more likely that different types of learning can be done independently.
A further insight into category learning that comes from a cognitive neuroscience perspective is the notion that there are multiple forms of category learning that depend on different brain systems. Although there may be fundamental similarities between different types of category learning at the behavioral level, they can be dissociated from each other in neuropsychological patients. For example, patients with Parkinson's disease perform normally in learning perceptual categories, such as in the dot pattern task described above, although they exhibit a profound impairment in the implicit learning of associations between cues and outcomes in the probabilistic classification task[16]. The hypothesis that these two types of category learning have different neural substrates is supported by neuroimaging data that demonstrate learning-related posterior cortical activity during performance of the dot pattern task [17], while the probabilistic classification task activates the caudate nucleus [18].
In real-life category-learning situations, multiple brain systems are
undoubtedly active. Exemplar-based knowledge, and thus the brain systems
underlying declarative memory, are probably also involved in category-learning
situations in which there are a limited number of salient category exemplars.
Rather than viewing category learning in terms of a dichotomy between
exemplar-based and category-knowledge based views, it might be more fruitful to
describe the circumstances under which different types of category learning
occur.
References
1. M.I. Posner and S.W. Keele, On the genesis of abstract ideas. J. Exp. Psychol. 77 (1968), pp. 353Æ363. Abstract-PsycINFO
2. E.H. Rosch, On the internal structure of perceptual and semantic categories. In: T.E. Moore Editor, Cognitive Development and the Acquisition of Language Academic Press (1973), pp. 111Æ144.
3. D.L. Hintzman, `Schema abstraction' in a multiple-trace memory model. Psychol. Rev. 95 (1986), pp. 528Æ551.
4. D.L. Medin and M.M. Schaffer, Context theory of classification learning. Psychol. Rev. 85 (1978), pp. 207Æ238. Abstract-PsycINFO
5. R.M. Nosofsky, Exemplar-based accounts of relations between classification, recognition, and typicality. J. Exp. Psychol. Learn. Mem. Cognit. 14 (1988), pp. 700Æ708. Abstract-PsycINFO
6. R.M. Nosofsky, Exemplar-based approach to relating categorization, indentification, and recognition. In: F.G. Ashby Editor, Multidimensional Models of Perception and Cognition Erlbaum (1992), pp. 363Æ393. Abstract-PsycINFO | Abstract-PsycINFO
7. B.J. Knowlton, S.J. Ramus and L.R. Squire, Intact artificial grammar learning in amnesia: dissociation of classification learning and explicit memory for specific instances. Psychol. Sci. 3 (1992), pp. 172Æ179. Abstract-PsycINFO | Abstract-PsycINFO
8. B.J. Knowlton and L.R. Squire, The learning of categories: parallel brain systems for item memory and category knowledge. Science 262 (1993), pp. 1747Æ1749. Abstract-PsycINFO | Abstract-EMBASE | Abstract-Elsevier BIOBASE
9. B.J. Knowlton, L.R. Squire and M.A. Gluck, Probabilistic classification learning in amnesia. Learn. Mem. 1 (1994), pp. 106Æ120. Abstract-PsycINFO | Abstract-Elsevier BIOBASE
10. R.M. Nosofsky and S.R. Zaki, Dissociations between categorization and recognition in amnesic and normal individuals: an exemplar-based interpretation. Psychol. Sci. 9 (1998), pp. 247Æ255. Abstract-PsycINFO | Full-text via CrossRef
11. S.B. Hamann and L.R. Squire, Intact perceptual memory in the absence of conscious memory. Behav. Neurosci. 111 (1997), pp. 850Æ854. Abstract-PsycINFO | Abstract-EMBASE
12. J.M. Reed et al., When amnesic patients perform well on recognition memory tests. Behav. Neurosci. 111 (1997), pp. 1163Æ1170. Abstract-PsycINFO | Abstract-EMBASE
13. L.R. Squire and B.J. Knowlton, Learning about categories in the absence of memory. Proc. Natl. Acad. Sci. U. S. A. 92 (1995), pp. 12470Æ12474. Abstract-BIOTECHNOBASE | Abstract-EMBASE | Abstract-Elsevier BIOBASE
14. B.J. Knowlton, J.A. Mangels and L.R. Squire, A neostriatal habit learning system in humans. Science 273 (1996), pp. 1399Æ1402. Abstract-PsycINFO | Abstract-PsycINFO | Abstract-Elsevier BIOBASE
15. P.J. Reber, B.J. Knowlton and L.R. Squire, Dissociable properties of memory systems: differences in the flexibility of declarative and nondeclarative knowledge. Behav. Neurosci. 110 (1996), pp. 861Æ871. Abstract-PsycINFO | Abstract-EMBASE
16. P.J. Reber and L.R. Squire, Intact learning of artificial grammars and intact category learning by patients with Parkinson's disease. Behav. Neurosci. (in press).
17. P.J. Reber, C.E.L. Stark and L.R. Squire, Cortical areas supporting category learning identified using functional MRI. Proc. Natl. Acad. Sci. U. S. A. 95 (1998), pp. 747Æ750. Abstract-BIOTECHNOBASE | Abstract-EMBASE | Full-text via CrossRef
18. R.A. Poldrack et
al., Striatal activation during acquisition of a cognitive skill. Neuropsychology
(in press).
tel: +1 310 825
5917 fax: +1 310 206 5895; email: knowlton@psych.ucla.edu
|
|||||||||||||||||||||||||||||||
|
|
Trends
in Cognitive Sciences |
||||||||||||||||||||||||||||||
|
|
|
|
|
|
|
|
Send feedback to ScienceDirect |
Your use of this service is governed by Terms and Conditions. Please review our Privacy Policy for details on how we protect information that
you supply.