When more is less: Negative exposure effects in unsupervised learning

In this article, two broad classes of models of unsupervised learning are compared: correlation tracking models, according to which learning is expected to increase monotonically with exposure to instances, and category invention models, which can accommodate specific violations of monotonicity (negative exposure effects). In two experiments, increasing the number of training instances had a negative rather than a positive effect on unsupervised learning, a clear violation of monotonicity. The results of these experiments are then compared with the predictions of two computational models, one a category invention model and the other a correlation tracking model. The category invention model was able to reproduce the qualitative pattern of results from the human data, whereas the correlation tracking model was not. Overall, these results provide strong evidence for the existence of a discrete category invention process in unsupervised learning.     

On the Interaction of Theory and Data in Concept Learning

Standard models of concept learning generally focus on deriving statistical properties of a category based on data (i.e., category members and the features that describe them) but fail to give appropriate weight to the contact between people's intuitive theories and these data. Two experiments explored the role of people's prior knowledge or intuitive theories on category learning by manipulating the labels associated with the category. Learning differed dramatically when categories of children's drawings were meaningfully labeled (e.g., “done by creative children”) compared to when they were labeled in a neutral manner. When categories are meaningfully labeled, people bring intuitive theories to the learning context. Learning then involves a process in which people search for evidence in the data that supports abstract features or hypotheses that have been activated by the intuitive theories. In contrast, when categories are labeled in a neutral manner, people search for simple features that distinguish one category from another. Importantly, the final study suggests that learning involves an interaction of people's intuitive theories with data, in which theories and data mutually influence each other. The results strongly suggest that straight-forward, relatively modular ways of incorporating prior knowledge into models of category learning are inadequate. More telling, the results suggest that standard models may have fundamental limitations. We outline a speculative model of learning in which the interaction of theory and data is tightly coupled. The article concludes by comparing the results to recent artificial intelligence systems that use prior knowledge during learning.     

How do people learn to allocate resources? Comparing two learning theories

How do people learn to allocate resources? To answer this question, 2 major learning models are compared, each incorporating different learning principles. One is a global search model, which assumes that allocations are made probabilistically on the basis of expectations formed through the entire history of past decisions. The 2nd is a local adaptation model, which assumes that allocations are made by comparing the present decision with the most successful decision up to that point, ignoring all other past decisions. In 2 studies, participants repeatedly allocated a capital resource to 3 financial assets. Substantial learning effects occurred, although the optimal allocation was often not found. From the calibrated models of Study 1, a priori predictions were derived and tested in Study 2. This generalization test shows that the local adaptation model provides a better account of learning in resource allocations than the global search model.     

Expanding the search for a linear separability constraint on category learning

Formal models of categorization make different predictions about the theoretical importance of linear separability. Prior research, most of which has failed to find support for a linear separability constraint on category learning, has been conducted using tasks that involve learning two categories with a small number of members. The present experiment used four categories with three or nine patterns per category that were either linearly separable or not linearly separable. With overall category structure equivalent across category types, the linearly separable categories were found to be easier to learn than the not linearly separable categories. An analysis of individual participants' data showed that there were more participants operating under a linear separability constraint when learning large categories than when learning small ones. Formal modeling showed that an exemplar model could not account for many of these data. These results are taken to support the existence of multiple processes in categorization.     

Blocking in category learning

Many theories of category learning assume that learning is driven by a need to minimize classification error. When there is no classification error, therefore, learning of individual features should be negligible. The authors tested this hypothesis by conducting three category-learning experiments adapted from an associative learning blocking paradigm. Contrary to an error-driven account of learning, participants learned a wide range of information when they learned about categories, and blocking effects were difficult to obtain. Conversely, when participants learned to predict an outcome in a task with the same formal structure and materials, blocking effects were robust and followed the predictions of error-driven learning. The authors discuss their findings in relation to models of category learning and the usefulness of category knowledge in the environment.     

Learning myopia: an adaptive recency effect in category learning

Recency effects (REs) have been well established in memory and probability learning paradigms but have received little attention in category earning research. Extant categorization models predict REs to be unaffected by learning, whereas a functional interpretation of REs, suggested by results in other domains, predicts that people are able to learn sequential dependencies and incorporate this information into their responses. These contrasting predictions were tested in 2 experiments involving a classification task in which outcome sequences were autocorrelated. Experiment 1 showed that reliance on recent outcomes adapts to the structure of the task, in contrast to models' predictions. Experiment 2 provided constraints on how sequential information is learned and suggested possible extensions to current models to account for this learning.     

Combining exemplar-based category representations and connectionist learning rules.

Adaptive network and exemplar-similarity models were compared on their ability to predict category learning and transfer data. An exemplar-based network (Kruschke, 1990a, 1990b, 1992) that combines key aspects of both modeling approaches was also tested. The exemplar-based network incorporates an exemplar-based category representation in which exemplars become associated to categories through the same error-driven, interactive learning rules that are assumed in standard adaptive networks. Experiment 1, which partially replicated and extended the probabilistic classification learning paradigm of Gluck and Bower (1988a), demonstrated the importance of an error-driven learning rule. Experiment 2, which extended the classification learning paradigm of Medin and Schaffer (1978) that discriminated between exemplar and prototype models, demonstrated the importance of an exemplar-based category representation. Only the exemplar-based network accounted for all the major qualitative phenomena; it also achieved good quantitative predictions of the learning and transfer data in both experiments.     

A rule-plus-exception model for classifying objects in continuous-dimension spaces

The authors propose a rule-plus-exception (RULEX) model for how observers classify stimuli residing in continuous-dimension spaces. The model follows in the spirit of the discrete-dimension version of RULEX developed by Nosofsky, Palmeri, and McKinley (1994). According to the model, observers learn categories by forming simple logical rules along single dimensions and by remembering occasional exceptions to those rules. In the continuous-dimension version of RULEX, the rules are formalized in terms of linear decision boundaries that are orthogonal to the coordinate axes of the psychological space. In addition, a similarity-comparison process governs whether stored exceptions are used to classify an object. The model provides excellent quantitative fits both to averaged classification transfer data and to distributions of generalizations observed at the individual-participant level. The modeling analyses suggest that, when multiple rules are available for solving a problem, averaged classification data often represent a probabilistic mixture of idiosyncratic rule-plus-exception strategies.     

A category-overshadowing effect in pigeons: support for the Common Elements Model of object categorization learning

A model proposing error-driven learning of associations between representations of stimulus properties and responses can account for many findings in the literature on object categorization by nonhuman animals. Furthermore, the model generates predictions that have been confirmed in both pigeons and people, suggesting that these learning processes are widespread across distantly related species. The present work reports evidence of a category-overshadowing effect in pigeons' categorization of natural objects, a novel behavioral phenomenon predicted by the model. Object categorization learning was impaired when a second category of objects provided redundant information about correct responses. The same impairment was not observed when single objects provided redundant information, but the category to which they belonged was uninformative, suggesting that this effect is different from simple overshadowing, arising from competition among stimulus categories rather than individual stimuli during learning.     

知道与不知道要学多少：类别学习中样例量的预期作用

采用“5/4模型”类别结构探讨了类别学习中样例量的预期作用。设置了两种学习条件(“知道样例量”和“不知道样例量”), 分别探讨两种学习条件下的学习效率、学习策略以及所形成的类别表征。106名大学生参加了实验, 结果表明：在类别学习中, 样例量的预期作用显著, 知道样例量组的学习效率高于不知道样例量组; 样例量的预期作用对类别学习效率的影响是通过影响学习过程中使用的策略来实现的; 样例量的预期作用不影响两种学习条件的学习后形成的类别表征, 且两种学习条件的被试自始至终表现出样例学习的表征模式。     

Test of a potential link between analytic and nonanalytic category learning and automatic, effortful processing

The link between automatic and effortful processing and nonanalytic and analytic category learning was evaluated in a sample of 29 college undergraduates using declarative memory, semantic category search, and pseudoword categorization tasks. Automatic and effortful processing measures were hypothesized to be associated with nonanalytic and analytic categorization, respectively. Results suggested that contrary to prediction strong criterion-attribute (analytic) responding on the pseudoword categorization task was associated with strong automatic, implicit memory encoding of frequency-of-occurrence information. Data are discussed in terms of the possibility that criterion-attribute category knowledge, once established, may be expressed with few attentional resources. The data indicate that attention resource requirements, even for the same stimuli and task, vary depending on the category rule system utilized. Also, the automaticity emerging from familiarity with analytic category exemplars is very different from the automaticity arising from extensive practice on a semantic category search task. The data do not support any simple mapping of analytic and nonanalytic forms of category learning onto the automatic and effortful processing dichotomy and challenge simple models of brain asymmetries for such procedures.     

A hybrid model of categorization

Category learning is often modeled as either an exemplar-based or a rule-based process. This paper shows that both strategies can be combined in a cognitive architecture that was developed to model other task domains. Variations on the exemplar-based random walk (EBRW) model of Nosofsky and Palmeri (1997b) and the rule-plus-exception (RULEX) rule-based model of Nosofsky, Palmeri, and McKinley (1994) were implemented in the ACT-R cognitive architecture. The architecture allows the two strategies to be mixed to produce classification behavior. The combined system reproduces latency, learning, and generalization data from three category-learning experiments--Nosofsky and Palmeri (1997b), Nosofsky et al., and Erickson and Kruschke (1998). It is concluded that EBRW and ACT-R have different but equivalent means of incorporating similarity and practice. In addition, ACT-R brings a theory of strategy selection that enables the exemplar and the rule-based strategies to be mixed.     

A knowledge-resonance (KRES) model of category learning

This article introduces a connectionist model of category learning that takes into account the prior knowledge that people bring to new learning situations. In contrast to connectionist learning models that assume a feedforward network and learn by the delta rule or backpropagation, this model, the knowledge-resonance model, or KRES, employs a recurrent network with bidirectional symmetric connection whose weights are updated according to a contrastive Hebbian learning rule. We demonstrate that when prior knowledge is represented in the network, KRES accounts for a considerable range of empirical results regarding the effects of prior knowledge on category learning, including (1) the accelerated learning that occurs in the presence of knowledge, (2) the better learning in the presence of knowledge of category features that are not related to prior knowledge, (3) the reinterpretation of features with ambiguous interpretations in light of error-corrective feedback, and (4) the unlearning of prior knowledge when that knowledge is inappropriate in the context of a particular category.     

Observed attention allocation processes in category learning

In two empirical studies of attention allocation during category learning, we investigate the idea that category learners learn to allocate attention optimally across stimulus dimensions. We argue that “optimal” patterns of attention allocation are model or process specific, that human learners do not always optimize attention, and that one reason they fail to do so is that under certain conditions the cost of information retrieval or use may affect the attentional strategy adopted by learners. We empirically investigate these issues using a computer interface incorporating an “information-board” display that collects detailed information on participants' patterns of attention allocation and information search during learning trials. Experiment 1 investigated the effects on attention allocation of distributing perfectly diagnostic features across stimulus dimensions versus within one dimension. The overall pattern of viewing times supported the optimal attention allocation hypothesis, but a more detailed analysis produced evidence of instance- or category-specific attention allocation, a phenomenon not predicted by prominent computational models of category learning. Experiment 2 investigated the strategies adopted by category learners encountering redundant perfectly predictive cues. Here, the majority of participants learned to distribute attention optimally in a cost–benefit sense, allocating attention primarily to only one of the two perfectly predictive dimensions. These results suggest that learners may take situational costs and benefits into account, and they present challenges for computational models of learning that allocate attention by weighting stimulus dimensions.     

Observed attention allocation processes in category learning

In two empirical studies of attention allocation during category learning, we investigate the idea that category learners learn to allocate attention optimally across stimulus dimensions. We argue that "optimal" patterns of attention allocation are model or process specific, that human learners do not always optimize attention, and that one reason they fail to do so is that under certain conditions the cost of information retrieval or use may affect the attentional strategy adopted by learners. We empirically investigate these issues using a computer interface incorporating an "information-board" display that collects detailed information on participants' patterns of attention allocation and information search during learning trials. Experiment 1 investigated the effects on attention allocation of distributing perfectly diagnostic features across stimulus dimensions versus within one dimension. The overall pattern of viewing times supported the optimal attention allocation hypothesis, but a more detailed analysis produced evidence of instance- or category-specific attention allocation, a phenomenon not predicted by prominent computational models of category learning. Experiment 2 investigated the strategies adopted by category learners encountering redundant perfectly predictive cues. Here, the majority of participants learned to distribute attention optimally in a cost-benefit sense, allocating attention primarily to only one of the two perfectly predictive dimensions. These results suggest that learners may take situational costs and benefits into account, and they present challenges for computational models of learning that allocate attention by weighting stimulus dimensions.     

Features of graded category structure: Generalizing the family resemblance and polymorphous concept models

Many real-world categories contain graded structure: certain category members are rated as more typical or representative of the category than others. Research has shown that this graded structure can be well predicted by the degree of commonality across the feature sets of category members. We demonstrate that two prominent feature-based models of graded structure, the family resemblance (Rosch & Mervis, 1975) and polymorphous concept models (Hampton, 1979), can be generalized via the contrast model (Tversky, 1977) to include both common and distinctive feature information, and apply the models to the prediction of typicality in 11 semantic categories. The results indicate that both types of feature information play a role in the prediction of typicality, with common features weighted more heavily for within-category predictions, and distinctive features weighted more heavily for contrast-category predictions. The same pattern of results was found in additional analyses employing rated goodness and exemplar generation frequency. It is suggested that these findings provide insight into the processes underlying category formation and representation.     

Gaze and informativeness during category learning: Evidence for an inverse relation

In what follows, we explore the general relationship between eye gaze during a category learning task and the information conveyed by each member of the learned category. To understand the nature of this relationship empirically, we used eye tracking during a novel object classification paradigm. Results suggest that the average fixation time per object during learning is inversely proportional to the amount of information that object conveys about its category. This inverse relationship may seem counterintuitive; however, objects that have a high-information value are inherently more representative of their category. Therefore, their generality captures the essence of the category structure relative to less representative objects. As such, it takes relatively less time to process these objects than their less informative companions. We use a general information measure referred to as representational information theory (Vigo, 2011a, 2013a) to articulate and interpret the results from our experiment and compare its predictions to those of three models of prototypicality.     

Problem structure and the use of base-rate information from experience

This article is concerned with the use of base-rate information that is derived from experience in classifying examples of a category. The basic task involved simulated medical decision making in which participants learned to diagnose hypothetical diseases on the basis of symptom information. Alternative diseases differed in their relative frequency or base rates of occurrence. In five experiments initial learning was followed by a series of transfer tests designed to index the use of base-rate information. On these tests, patterns of symptoms were presented that suggested more than one disease and were therefore ambiguous. The alternative or candidate diseases on such tests could differ in their relative frequency of occurrence during learning. For example, a symptom might be presented that had appeared with both a relatively common and a relatively rare disease. If participants are using base-rate information appropriately (according to Bayes' theorem), then they should be more likely to predict that the common disease is present than that the rare disease is present on such ambiguous tests. Current classification models differ in their predictions concerning the use of base-rate information. For example, most prototype models imply an insensitivity to base-rate information, whereas many exemplar-based classification models predict appropriate use of base-rate information. The results reveal a consistent but complex pattern. Depending on the category structure and the nature of the ambiguous tests, participants use base-rate information appropriately, ignore base-rate information, or use base-rate information inappropriately (predict that the rare disease is more likely to be present). To our knowledge, no current categorization model predicts this pattern of results. To account for these results, a new model is described incorporating the ideas of property or symptom competition and context-sensitive retrieval.     

A rational analysis of rule-based concept learning

This article proposes a new model of human concept learning that provides a rational analysis of learning feature-based concepts. This model is built upon Bayesian inference for a grammatically structured hypothesis space—a concept language of logical rules. This article compares the model predictions to human generalization judgments in several well-known category learning experiments, and finds good agreement for both average and individual participant generalizations. This article further investigates judgments for a broad set of 7-feature concepts—a more natural setting in several ways—and again finds that the model explains human performance.