Dorrit Billman

Event Category Learning

This research investigated the learning of event categories, in particular, categories of simple animated events, each involving a causal interaction between 2 characters. Four experiments examined whether correlations among attributes of events are easier to learn when they form part of a rich correlational structure than when they are independent of other correlations. Event attributes (e.g., state change, path of motion) were chosen to reflect distinctions made by verbs. Participants were presented with an unsupervised learning task and were then tested on whether the organization of correlations affected learning. Correlations forming part of a system of correlations were found to be better learned than isolated correlations. This finding of facilitation from correlational structure is explained in terms of a model that generates internal feedback to adjust the salience of attributes. These experiments also provide evidence regarding the role of object information in events, suggesting that this role is mediated by object category representations.

Systems of correlations in rule and category learning: Use of structured input in learning syntactic categories

Abstract Three studies with artificial grammars investigated one property of natural language and one learning procedure which capitalises on this property to aid learning. Learning syntax is difficult because the rules and categories of syntax form a mutually defining system. Structural analogies to other systems (e.g. semantics, interaction) are limited and unreliable. On the other hand, treating syntax as a formal system and applying distributional analysis poses an extremely difficult learning problem. The property investigated is systematicity, coherence among multiple interpredictive features marking syntactic categories. The learning procedure is a method for directing attention to predictive features. This procedure predicts that an individual correlational rule or pattern will be learned more easily when it is part of a system of rules among intercorrelated features than when the identical rule occurs in isolation—complexity facilitates learning. The artificial grammar experiments varied the stru...

Induction from a Single Instance: Formation of a Novel Category.

This study examines whether preschoolers can use information from a known category to induce a characteristic attribute of a novel, contrasting category based on a single instance. We showed 32 four-year-olds three instances of a Given Category and one instance of a Target Category. These objects could vary along two attribute dimensions, such as color and shape. All instances of the Given Category shared identical values of one attribute (e.g., all were blue), but could have different values of the other attribute (e.g., a circle, a square, and a triangle). The single instance of the Target Category was different from the Given on both attribute dimensions (e.g., a red diamond). Children gave yes/no judgements as to whether additional objects were instances of the Target Category. There were two possible sources of information about the relevance of an attribute to classification: explicit (labeling) and implicit (variation in the Given Category). There were four conditions such that each source of information was either available or not. Both types of information were effective in eliciting inductions of the relevant kind of attribute and the characteristic value of this attribute in the novel category (explicit: p = .0004; implicit: p = .031). This suggests that children use an inductive bias that the instances of two related but distinct categories tend to be alike in the same way.

Acquiring and Combining Overlapping Concepts

This article presentsOloc, an incremental concept formation system that learns and uses overlapping concepts.Oloc learns probabilistic concepts that have overlapping extensions and does so to maximize expected predictive accuracy. When making predictions,Oloc can combine multiple overlapping concepts.

Consistent contrast aids concept learning.

We suggest that coherence among concepts and correspondence between concepts and the world are important in concept learning. We identify one aspect of coherence, consistent contrast, and investigate its role in supervised concept learning. Concepts that contrast consistently carry information about the same attributes across the concepts within a contrast set. Concepts that contrast inconsistently predict and are predicted by values of different attributes. Experiment 1 revealed a large advantage for consistent contrast in learning and generalization. Experiment2 pitted similarity against consistency and still revealed an advantage of consistency. Experiment 2 also broadened the range of tasks considered to include inductions about novel categories and subjects’ category descriptions. We discuss relations to theories of concept learning, to attentional mechanism, and to alignability, and we suggest practical implications.

Needs Analysis and Technology Evaluation: Evaluation Case Study of Alternative Software for Controller Planning Work—Part 2

A NASA Mission Control group provided us the opportunity to conduct a needs analysis, produce software guided by the needs analysis, and evaluate the software prototype. This paper reports our discriminative evaluation of new prototype software for Attitude Determination and Control Officers, who plan and execute maneuvers of the International Space Station with their Russian counterparts. On a specific, pragmatic level, our evaluation provided evidence of large performance improvement. Relative to legacy software, the new software reduced time and errors by half in a laboratory experiment using valid tasks identified from the needs analysis. Our discriminative evaluation also isolated specific benefits attributable to specific improvements in alignment of technology to the work. Our discriminative evaluation isolated contributing factors affecting performance by using item sets designed to differentially engage software components. This approach for identifying specific factors contributing to performance improvement may be reusable in many situations, particularly where it is not feasible to develop and test many software versions, each differing in just a single factor of interest.

Representational specificity and concept learning

Publisher Summary This chapter discusses the representational specificity and concept learning. When a young child is fed by an adult, at first she assumes that all generally similar creatures, that is, all adults, would treat her equally well. With more experience, she comes to expect food from a smaller group of adults, such as all familiar adults, and then from a yet more specific class of creatures, such as mother, father, and grandmother (MFG). On the other hand, the child could initially expect feeding from a very specific set, mother only, and with experience generalize to female relatives (MG) and then to larger classes, such as MFG. Clearly, both approaches, starting with general classes or starting with specific classes, can produce appropriate results. Both converge toward an optimal level of generality. Even though the two approaches attain the same result, one could be more efficient. For instance, if one approach generally requires fewer examples before converging on the correct class, that method could be used in machine learning algorithms and would be expected to be used in natural systems.

11 Modeling Category Learning and Use: Representation and Processing

Publisher Summary This chapter discusses two ways to investigate general classes of concept and categorization models, given the tradeoff and indeterminacy between claims about the nature of processing and of representation. One approach is to move away from the mechanism language of representation and processing to test classes of models. In this approach, instead of detailing a mechanism, the effort is to precisely specify a class of mapping functions that the representation and process—whatever they may be—jointly compute. In the case of category learning, the function domain is a set of learning instances plus the target items to categorize, and the function range is the set of possible category judgments. Thus, any such function maps from the learning experience plus current example onto a categorization judgment of the current example. Testing a class of models, then, consists of testing a specified class of functions. Some property inconsistent with the specified class of mapping functions must be identified and the property can then be tested. Any test showing the necessity of the critical property is then a demonstration of the inadequacy of the specified class.

Critique of structural analysis in modeling cognition: A case study of Jackendoff's theory

Abstract Modeling cognition by structural analysis of representation leads to systematic difficulties which are not resolvable. We analyse the merits and limits of a representation‐based methodology to modeling cognition by treating Jackendoffs Consciousness and the Computational Mind as a good case study. We note the effects this choice of methodology has on the view of consciousness he proposes, as well as a more detailed consideration of the computational mind. The fundamental difficulty we identify is the conflict between the desire for modular processors which map directly onto representations and the need for dynamically interacting control. Our analysis of this approach to modeling cognition is primarily directed at separating merits from problems and inconsistencies by a critique internal to this approach; we also step outside the framework to note the issues it ignores.