Neural Net Modeling and Developmental Psychology
Saturday, 8:30 AM - 10:15 AM
Hynes CC 310

Chair: Jeffrey Elman, University of California, San Diego

Biographical Sketch
Jeffrey Elman received his Bachelor’s degree from Harvard in 1969 and his Ph.D. in Linguistics from the University of Texas at Austin in 1977. That same year, he joined the faculty at UCSD, where he has remained ever since. In 1986, he helped found the Department of Cognitive Science, where he served as Chair from 1994 to 1998. Elman is currently Chancellor’s Associates’ Distinguished Professor of Cognitive Science, Acting Dean of the Division of Social Sciences at UCSD, and Founding Co-Director of the Kavli Institute for Brain and Mind. He is the 2007 recipient of the David E. Rumelhart Prize for Contributions to the Theoretical Foundations of Cognitive Science.

Abstract
In recent years, a new computational framework has emerged for modeling developmental phenomena: Connectionism (or neural network models). This paradigm has excited great interest and a certain amount of controversy. Among other things, neural network models provide an alternative to more traditional symbolic accounts of development, cognition, and neuropsychology. Many people feel that the connectionist paradigm has led to new insights into old problems. Importantly, it has also spurred new experiments that are aimed at testing predictions made by models. This symposium has two goals. The first goal is to provide a tutorial introduction to neural network modeling in a way that is accessible to non-modelers and focuses on central principles and assumptions. The second goal is to present results of recent modeling work that bear on topics in developmental psychology, including semantic cognition, the notion of domain specificity and modularity, infant perception and categorization, developmental disorders and atypical development, and language acquisition.

Presentation 1
Neural Net Modeling and Developmental Psychology: An Overview
Speaker: Jeffrey Elman, University of California, San Diego

Presentation 2
Modeling Infant Perception
Speaker: Denis Mareschal, University of London

Biographical Sketch
Denis Mareschal obtained his first degree in Physics and Theoretical Physics from Cambridge University. He then completed a Masters in Psychology from McGill University before moving on to complete a PhD at Oxford University. He has received the Marr prize from the Cognitive Science Society (USA), the Young Investigator Award from the International Society on Infant Studies (USA), and the Margaret Donaldson Prize from the British Psychological Society. His research centers on developing mechanistic models of perceptual and cognitive development in infancy and childhood. He has three children and is currently Professor at Birkbeck College, University of London.

Abstract
Habituation and visual preference techniques are central to the way that we explore infant perceptual and cognitive development. Of course, this begs the question of what exactly is happening during familiarization. Connectionist models provide a tool for exploring how infants develop internal representation during habituation. I will argue that having a process model of HOW learning takes places raises different kinds of empirical questions. As an example, I will focus on recent work exploring early infant perceptual categorization. While young infants seem to form categories similar to the corresponding adult concepts, their categories sometimes show idiosyncratic properties. For example, 3-month-olds will form a category of CAT that excludes dogs but also a category of DOG that excludes cats. I will show how developing a connectionist model that learns to form categories on the basis the same photographs of cats and dogs used to test infants reveals why infants form such asymmetric categories. The models make subtle predictions about how manipulating the object images will change the infants' behavior and their apparent understanding these categories. This exchange of modeling and empirical work illustrates the necessary dialogue between developing formal theories of learning and development, and rigorous empirical testing.

Presentation 3
Neural Network Approaches to Developmental Disorders: Modeling Cognitive Variability
Speaker: Michael Thomas, University of London

Biographical Sketch
Michael Thomas completed a BSc in Psychology at the University of Exeter and an MSc in Cognitive Science at the University of Birmingham in the UK. He then completed a D.Phil at the University of Oxford on behavioral and computational studies of bilingualism. He worked with Annette Karmiloff-Smith at the Institute in Child Health in London investigating neurogenetic disorders, before moving to Birkbeck College, University of London, where he is now Reader in Cognitive Neuropsychology. His research focuses on language and cognitive development, and in particular, empirical and computational approaches to understanding the cognitive variability seen in typical children and in children with developmental disorders (see www.psyc.bbk.ac.uk/research/DNL/).

Abstract
Neural network models have provided an increasing body of work on the mechanisms that drive normal cognitive development. These models usually attempt to capture the developmental trajectory exhibited by the average child, identified from group-averaged data. However, cognitive development shows variability both within the normal range (in terms of intelligence) and outside the normal range (in terms of developmental disorders). In this talk, I review how neural network models have been applied to the study of behavioral deficits found in disorders like autism, dyslexia, Specific Language Impairment, and Attention Deficit Hyperactivity Disorder. These models encourage researchers to place the developmental process itself at the heart of their explanations of impairments, and incorporate concepts such as compensation and redundancy. I briefly describe how these models link to methods of clinical intervention. I then indicate how neural network models can also shed light on possible mechanisms underlying variability within the normal range as part of research into intelligence. Lastly I close by drawing links between explanations of typical and atypical variability and recent approaches to studying the influence of genes on human behaviour.

Presentation 4
What Is Domain-Specific About Cognitive Development? A Connectionist Perspective
Speaker: James L. McClelland, Stanford University

Biographical Sketch
James L. (Jay) McClelland is Professor of Psychology and Director of the Center for Mind, Brain and Computation at Stanford. At UCSD in the 1980's, McClelland and David Rumelhart led the group producing the 2-volume work Parallel Distributed Processing. McClelland has subsequently investigated development, learning, and memory, introducing a Complementary Learning Systems theory of the brain organization of learning with Bruce McNaughton and Randy O'Reilly and a Parallel Distributed Processing approach to the development and disintegration of semantic knowledge with Tim Rogers. Among other honors, McClelland is an APS William James Fellow and a member of the National Academy of Sciences.

Abstract
Connectionist models offer general principles of processing, representation and learning. In my own work, I have adopted the approach of seeking to use these general principles as the basis for explaining domain specific facts about cognitive and linguistic development. In the course of this work, my collaborators and I have come to articulate an overall architecture for learning and development that is domain-general in many ways but has an articulated structure, including interdependent components that play relatively specialized roles and may have initial predispositions that are shaped by pre-natal developmental processes. In my talk I will describe the general principles, the overall architecture for learning and development and the ways in which the specialized components reflect the domain-general principles. I will also discuss how domain-specific constraints on processing can arise from domain general principles of learning applied to domain-specific experience.