Robert M. Nosofsky - US grants

This research will study how people learn to classify objects and how categories are represented in memory. The research will be guided by the continued development and testing of a mathematical model, the exemplar model, formalizing the view that classification decisions are based on similarity comparisons between presented objects and stored exemplars. The research will involve the study of performance relations between categorization and other choice tasks people are often asked to perform, including identification and recognition. In the categorization task, people classify items into groups; in the identification task, each item is assigned a unique response; in the recognition task, people simply judge whether items are old or new. The exemplar model and another class of model, rule- based classification models, make different predictions about the relationships among results of experiments involving the three choice tasks. Therefore, the experiments should help to distinguish between these two kinds of theories. Theoretical research accompanying the experimentation will be aimed at elucidating the nature of stimulus bias and asymmetric similarity relations. The process of categorization is among the most fundamental of mental activities. It allows people to bring order and organization to their environment, and it is a building block of more complex cognitive activities such as reasoning, thinking, and problem solving. A clear understanding of the nature of categorization processes will be instrumental in advancing our understanding of human thought.

The long-term objective of this research is the development of a general formal model of perceptual categorization and memory, which interrelates performance across a variety of tasks, including classification, identification, and recognition. The present project is organized around the continued development and testing of Nosofsky's (1986, 1987) generalized context model (GCM), which is a highly successful mathematical model of perceptual categorization and recognition. According to the GCM, people represent categories by storing individual exemplars in memory, and make classification and recognition decisions on the basis of similarity comparisons with the stored exemplars. Similarity-scaling techniques are used to represent sets of exemplars in multidimensional psychological spaces. These derived spaces are then used in conjunction with the formal model to quantitatively predict performance in a variety of independent tasks. Although highly successful to date, most previous tests of the GCM have occurred in highly simplified perceptual domains that allowed one to maintain precise control over the fundamental variables of interest. One aim of the present work is to test the model in a much richer, complex domain than has thus far been attempted, and demonstrate its applicability using "ill-defined," natural category structures. A second aim involves the development and testing of dynamic versions of the model, that should allow it to characterize processes of classification learning and changes in category representations as a function of experience. The project involves a continuing interplay among theory development, experimental testing, and modification of theory in line with newly obtained empirical results. Understanding the fundamental processes of categorization and recognition is one of the central goals of research in memory and cognition. Some direct health-related applications of the present work would include providing information about how radiologists make disease classifications on the basis of imperfect information provided in X-ray displays, and constructing mental illness classifications on the basis of reported symptomatology.

The long-term objective of this research is the development of a computational model of perceptual categorization and memory, which interrelates performance across a variety of tasks, including classification, identification, old-new recognition, and same- different judgment. The present project is organized around the continued development and testing of Nosofsky and Palmeri's exemplar-based random walk (EBRW) model. According to the EBRW, people represent categories by storing individual exemplars in memory. Test objects cause individual exemplars to be retrieved based on how similar the objects are to the exemplars. The retrieved exemplars provide evidence that enters into a random-walk process for making classification decisions. The EBRW goes beyond previous work by providing a detailed processing account of the time course of categorization decision making, thereby allowing the model to jointly predict classification choice probabilities and response times. One goal of the new work is to extend the EBRW with a stochastic dimensional encoding process to allow it to predict response times for separable-dimension stimuli as well as integral- dimension ones. A second goal is to extend the model to the domain of multidimensional same-different judgment. Finally, the project will investigate the extent to which the exemplar-based model can account for a wide variety of empirical phenomena which previous investigators have recently interpreted in terms of rule abstraction or prototype formation. Understanding the fundamental processes of perceptual categorization and recognition is one of the central goals of research in memory and cognition. A direct health-related application of the present work would be to provide information about how radiologists make disease classifications on the basis of imperfect information contained in X-ray displays, with the ultimate goal of developing training techniques as well as computer technology to assist in radiological decision making.

A ubiquitous component of science education is learning the key categories of each target domain. This project, a multidisciplinary collaboration of cognitive scientists, geologists, and science education researchers at Indiana University and Washington University, will take basic research findings on human learning and attempt to develop diagnostic tools that can be matched with instructional technique to facilitate the learning of scientific classifications. Rock categorization will be used as the example target domain because it provides challenges that are representative of scientific classification learning more generally, but the training insights that are generated should be applicable across multiple scientific domains. A critical practical issue in education research concerns how to explore the vast space of possible instructional variations. A key advantage of the project is that the researchers will delimit that space by testing specific hypotheses, drawn from successful formal models of human classification learning combined with principles from the training literature, about how content should be delivered to optimize learning of scientific classifications. The researchers will derive novice and expert representations of rock classifications, including the dimensions they attend to. This work bridges laboratory-based mathematical modeling research with more applied research: Instruction using real rocks in authentic learning situations will be contrasted with instruction delivered over computers. Principles that the researchers discover in comparisons of experts and novices should be useful in the development of diagnostic tools for future applications in the classroom and the field. The project fits centrally into the EHR Core Research (ECR) program goal of conducted funamental research and building enduring research foundations for STEM learning.

The studies will entail fundamental scaling work to derive psychological similarity representations for the rock stimuli. Derivation of these representations is a prerequisite for rigorous application of the models of classification that will guide the subsequent empirical training studies. These representations will also provide important insights concerning the major psychological dimensions along which the rock stimuli are organized as well as how the rock category distributions are configured in the multidimensional similarity space. These student representations will be contrasted with those derived from expert geologists. It is highly likely that the experts will have learned to focus attention on dimensions that are far mor ediagnostic than those used by the students. Empirical investigations of these different multidimensional solutions should yield important information regarding fundamental parameters for how most efficiently to support students' learning of the rock categories. These include identifying: i) the optimal training instances to support learning and generalization, ii) the optimal sequencing of these training instances, and iii) the preferred training density for particular subtypes of hierarchically organized category distributions.