Learning about the internal structure of categories through classification and feature inference

Previous research on category learning has found that classification tasks produce representations that are skewed toward diagnostic feature dimensions, whereas feature inference tasks lead to richer representations of within-category structure. Yet, prior studies often measure category knowledge through tasks that involve identifying only the typical features of a category. This neglects an important aspect of a category's internal structure: how typical and atypical features are distributed within a category. The present experiments tested the hypothesis that inference learning results in richer knowledge of internal category structure than classification learning. We introduced several new measures to probe learners' representations of within-category structure. Experiment 1 found that participants in the inference condition learned and used a wider range of feature dimensions than classification learners. Classification learners, however, were more sensitive to the presence of atypical features within categories. Experiment 2 provided converging evidence that classification learners were more likely to incorporate atypical features into their representations. Inference learners were less likely to encode atypical category features, even in a “partial inference” condition that focused learners' attention on the feature dimensions relevant to classification. Overall, these results are contrary to the hypothesis that inference learning produces superior knowledge of within-category structure. Although inference learning promoted representations that included a broad range of category-typical features, classification learning promoted greater sensitivity to the distribution of typical and atypical features within categories.     

Learning and retention through predictive inference and classification

Work in category learning addresses how humans acquire knowledge and, thus, should inform classroom practices. In two experiments, we apply and evaluate intuitions garnered from laboratory-based research in category learning to learning tasks situated in an educational context. In Experiment 1, learning through predictive inference and classification were compared for fifth-grade students using class-related materials. Making inferences about properties of category members and receiving feedback led to the acquisition of both queried (i.e., tested) properties and nonqueried properties that were correlated with a queried property (e.g., even if not queried, students learned about a species' habitat because it correlated with a queried property, like the species' size). In contrast, classifying items according to their species and receiving feedback led to knowledge of only the property most diagnostic of category membership. After multiple-day delay, the fifth-graders who learned through inference selectively retained information about the queried properties, and the fifth-graders who learned through classification retained information about the diagnostic property, indicating a role for explicit evaluation in establishing memories. Overall, inference learning resulted in fewer errors, better retention, and more liking of the categories than did classification learning. Experiment 2 revealed that querying a property only a few times was enough to manifest the full benefits of inference learning in undergraduate students. These results suggest that classroom teaching should emphasize reasoning from the category to multiple properties rather than from a set of properties to the category.     

Learning phonetic categories by tracking movements

We explore in this study how infants may derive phonetic categories from adult input that are highly variable. Neural networks in the form of self-organizing maps (SOMs; ) were used to simulate unsupervised learning of Mandarin tones. In Simulation 1, we trained the SOMs with syllable-sized continuous F(0) contours, produced by multiple speakers in connected speech, and with the corresponding velocity profiles (D1). No attempt was made to reduce the large amount of variability in the input or to add to the input any abstract features such as height and slope of the F(0) contours. In the testing phase, reasonably high categorization rate was achieved with F(0) profiles, but D1 profiles yielded almost perfect categorization of the four tones. Close inspection of the learned prototypical D1 profile clusters revealed that they had effectively eliminated surface variability and directly reflected articulatory movements toward the underlying targets of the four tones as proposed by . Additional simulations indicated that a further learning step was possible through which D1 prototypes with one-to-one correspondence to the tones were derived from the prototype clusters learned in Simulation 1. Implications of these findings for theories of language acquisition, speech perception and speech production are discussed.     

Feature inference and the causal structure of categories

The purpose of this article was to establish how theoretical category knowledge-specifically, knowledge of the causal relations that link the features of categories-supports the ability to infer the presence of unobserved features. Our experiments were designed to test proposals that causal knowledge is represented psychologically as Bayesian networks. In five experiments we found that Bayes' nets generally predicted participants' feature inferences quite well. However, we also observed a pervasive violation of one of the defining principles of Bayes' nets-the causal Markov condition-because the presence of characteristic features invariably led participants to infer yet another characteristic feature. We argue that this effect arises from a domain-general bias to assume the presence of underlying mechanisms associated with the category. Specifically, people take an exemplar to be a "well functioning" category member when it has most or all of the category's characteristic features, and thus are likely to infer a characteristic value on an unobserved dimension.     

Learning time-varying categories

Many kinds of objects and events in our world have a strongly time-dependent quality. However, most theories about concepts and categories either are insensitive to variation over time or treat it as a nuisance factor that produces irrational order effects during learning. In this article, we present two category learning experiments in which we explored peoples’ ability to learn categories whose structure is strongly time-dependent. We suggest that order effects in categorization may in part reflect a sensitivity to changing environments, and that understanding dynamically changing concepts is an important part of developing a full account of human categorization.     

Compound cue processing within the fast and frugal heuristics approach in nonlinearly separable environments

Three experiments investigated whether participants used Take The Best (TTB) Configural, a fast and frugal heuristic that processes configurations of cues when making inferences concerning which of two alternatives has a higher criterion value. Participants were presented with a compound cue that was nonlinearly separable from its elements. The compound was highly valid in Experiments 1 and 2, but invalid in Experiment 3. Participants’ causal mental models were manipulated via instructions: participants were either told that cues acted through the same causal mechanism (configural causal model), through different causal mechanisms (elemental causal model), or the causal mechanisms were not specified (neutral causal model). A high percentage of participants used TTB-Configural when they had a configural causal model and a highly valid compound existed, suggesting that causal knowledge can be incorporated in otherwise very basic cognitive mechanisms to allow fine-grained adaptation to complex task structures.     

Knowledge-based classification of ill-defined categories

In this paper, I report an exploratory study which investigated the role that prior knowledge plays in influencing classification learning. Under neutral or knowledge-imposing instructions, subjects learned to classify exemplars into categories that either were or were not linearly separable. Linearly separable categories are those categories whose members can be correctly classified based on an additive summation of weighted attribute information. Following category learning, the subjects were given transfer tests. A major finding was that knowledge facilitated the learning of linearly separable categories but interfered with the learning of not linearly separable categories. Quantitative analyses revealed that the knowledge facilitated category learning of the linearly separable categories by influencing the storage and reliance on both prototypical and exemplar information.     

Classification versus inference learning contrasted with real-world categories

Categories are learned and used in a variety of ways, but the research focus has been on classification learning. Recent work contrasting classification with inference learning of categories found important later differences in category performance. However, theoretical accounts differ on whether this is due to an inherent difference between the tasks or to the implementation decisions. The inherent-difference explanation argues that inference learners focus on the internal structure of the categories—what each category is like—while classification learners focus on diagnostic information to predict category membership. In two experiments, using real-world categories and controlling for earlier methodological differences, inference learners learned more about what each category was like than did classification learners, as evidenced by higher performance on a novel classification test. These results suggest that there is an inherent difference between learning new categories by classifying an item versus inferring a feature.     

Inference and classification learning of abstract coherent categories

Category learning research has primarily focused on how people learn to classify items using simple observable features. However, classification is only 1 way to learn categories. In addition, many concepts have an underlying coherence that explains the featural similarity among exemplars, such as abstract coherent concepts whose instances differ greatly on their observable features. In 3 experiments, the authors investigated how abstract coherent categories are acquired through 2 common means of category learning, classification and inference. Because inference promotes more focus on within-category information than does classification, they hypothesized that inference learning would lead to a better understanding of the underlying coherence of abstract coherent categories. All 3 experiments support this prediction.     

Category representation for classification and feature inference

This research's purpose was to contrast the representations resulting from learning of the same categories by either classifying instances or inferring instance features. Prior inference learning research, particularly T. Yamauchi and A. B. Markman (1998), has suggested that feature inference learning fosters prototype representation, whereas classification learning encourages exemplar representation. Experiment 1 supported this hypothesis. Averaged and individual participant data from transfer after inference training were better fit by a prototype than by an exemplar model. However, Experiment 2, with contrasting inference learning conditions, indicated that the prototype model was mimicking a set of label-based bidirectional rules, as determined by the inference learning task demands in Experiment 1. Only the set of rules model accounted for all the inference learning conditions in these experiments.     

Learning abstract relations from using categories

When people learn categories, the importance of the features and relations in the category representation reflects both their diagnosticity for classification and their relevance to the use of the category. In earlier work in which the influence of category use on the representation has been shown, only cases in which the features and relations were simple, observable, and very specific were examined. Learners may begin to understand the underlying similarities of category members by using the categories. In the four experiments presented here, learners applied a simple category-specific formula to category members. The test results showed that the learners had incorporated relations among features from this use, including cases in which the relations were abstract. This learning occurred even though the relations were actually not predictive of category membership but just perceived to be so as a function of the use.     

Learning categories by touch: On the development of holistic and analytic processing

The development of holistic and analytic processing often studied in the visual domain was investigated in haptics. Children 3 to 9 years of age and adults had to categorize haptic exemplars that varied systematically in four attributes (size, shape, surface texture, and weight). The subjects could learn the categories either analytically--that is, by focusing on a single attribute--or holistically--that is, in terms of overall similarity. The data show that even the youngest children learned the haptic categories far more often in an analytic mode than in a holistic mode. Nevertheless, an age trend was observed, referring to the attributes that the analytic learners used for their categorization. The children preferred substance-related attributes, especially surface texture, whereas the adults preferred structure-related attributes, especially shape. Thus, it appears that analytic and/or holistic processing in category learning develops in a similar manner in the visual and haptic domains.     

Dimensional,similarity, and configurai classification of integral and separable stimuli

Subjects classified sets of two-dimensional stimuli into two or three groups in any way they wished. For integral stimuli, classification was based on numerical balance and the similarity structure. For separable stimuli, classification was based on the dimensional structure and the configurai properties of the stimulus set. These results replicate and extend those of Handel and Imai (1972). In addition, they suggest that no closed set of stimulus concepts is possible.     

Learning categories by making predictions: An investigation of indirect category learning

Categories are learned in many ways, but studies of category learning have generally focused on classification learning. This focus may limit the understanding of categorization processes. Two experiments were conducted in which participants learned categories of animals by predicting how much food each animal would eat. We refer to this as indirect category learning, because the task andthe feedback were not directly related to category membership, yet category learning was necessary for good performance in the task. In the first experiment, we compared the performance of participants who learned the categories indirectly with the performance of participants who first learned to classify the objects. In the second experiment, we replicated the basic findings and examined attention to different features during the learning task. In both experiments, participants who learned in the prediction-only condition displayed a broader distribution of attention than participants who learned in the classification-and-prediction condition did. Some participants in the prediction-only group learned the family resemblance structure of the categories, even when a perfect criterial attribute was present. In contrast, participants who first learned to classify the objects tended to learn the criterial attribute.     

Learning and hypothesis testing in probabilistic inference tasks

The relation between learning of probabilistic inference tasks and the availability of correct hypotheses for the tasks in individual subjects was investigated in two experiments. The results showed that having the correct hypotheses as measured before learning the tasks was neither a sufficient nor a necessary condition for learning the task. It was concluded that the hypothesis measurements are not reliable enough to allow predictions on an individual level. The results also showed that subjects have not only the four basic functions: positive linear, negative linear, U-shaped and inversely U-shaped, in their hypothesis hierarchies but also J-shaped functions. That the subjects can leam such functions is thus no discodinnation of the hypotheses sampling conception of learning as assumed in earlier studies.     

Better mood and better performance. Learning rule-described categories is enhanced by positive mood

Theories of mood and its effect on cognitive processing suggest that positive mood may allow for increased cognitive flexibility. This increased flexibility is associated with the prefrontal cortex and the anterior cingulate cortex, both of which play crucial roles in hypothesis testing and rule selection. Thus, cognitive tasks that rely on behaviors such as hypothesis testing and rule selection may benefit from positive mood, whereas tasks that do not rely on such behaviors should not be affected by positive mood. We explored this idea within a category-learning framework. Positive, neutral, and negative moods were induced in our subjects, and they learned either a rule-described or a non-rule-described category set. Subjects in the positive-mood condition performed better than subjects in the neutral- or negative-mood conditions in classifying stimuli from rule-described categories. Positive mood also affected the strategy of subjects who classified stimuli from non-rule-described categories.     

Effects of classification context on categorization in natural categories

The patterns of classification of borderline instances of eight common taxonomic categories were examined under three different instructional conditions to test two predictions: first, that lack of a specified context contributes to vagueness in categorization, and second, that altering the purpose of classification can lead to greater or lesser dependence on similarity in classification. The instructional conditions contrasted purely pragmatic with more technical/quasi-legal contexts as purposes for classification, and these were compared with a no-context control. The measures of category vagueness were between-subjects disagreement and within-subjects consistency, and the measures of similarity-based categorization were category breadth and the correlation of instance categorization probability with mean rated typicality, independently measured in a neutral context. Contrary to predictions, none of the measures of vagueness, reliability, category breadth, or correlation with typicality were generally affected by the instructional setting as a function of pragmatic versus technical purposes. Only one subcondition, in which a situational context was implied in addition to a purposive context, produced a significant change in categorization. Further experiments demonstrated that the effect of context was not increased when participants talked their way through the task, and that a technical context did not elicit more all-or-none categorization than did a pragmatic context. These findings place an important boundary condition on the effects of instructional context on conceptual categorization.     

Learning complex rules in probabilistic inference tasks

Abstract.— A hypothesis-sampling theory for rule learning predicts that only those rules that are available for sampling can be learned. Earlier results show that subjects have a very limited set of hypotheses about rules relating scaled cue and criterion variables, consisting mainly of linear and symmetric quadratic functions, but not of complex functions, such as J-shaped rules. Tasks requiring the use of such rules should, therefore, not be learned. The results of the present experiment show, however, that subjects are able to find J-shaped relations. The results were interpreted to mean that subjects are able to construct hypotheses, and that they are not limited to sampling of hypotheses from a preestablished set of possible hypotheses.     

Working memory supports inference learning just like classification learning

Recent research has found a positive relationship between people's working memory capacity (WMC) and their speed of category learning. To date, only classification-learning tasks have been considered, in which people learn to assign category labels to objects. It is unknown whether learning to make inferences about category features might also be related to WMC. We report data from a study in which 119 participants undertook classification learning and inference learning, and completed a series of WMC tasks. Working memory capacity was positively related to people's classification and inference learning performance.