No Title

**NIPS*95 TUTORIAL PROGRAM**

November 27, 1995

Abstracts

Session I: 9:30--11:30 a.m.

"Functional Anatomy of Primate Vision"
Gary Blasdel, Harvard Medical School

This tutorial will deal with the visual system of primates as a sequence of two dimensional, nonlinear transforms, whose cumulative effect is to abstract vital information about the world shaped by the reflectance properties of objects in the surrounding space. It will emphasize basic information about the anatomical and physiological organization of neuronal elements at every stage, as well as current insights into the visual processing strategies that are thought to occur with every change. The topics covered will include: 1. Functional anatomies of the retina, LGN, striate and extra-striate cortex. 2. Rules governing the representation of visual information in every layer at every stage, and rules governing lateral interactions that modify these representations. 3. Rules governing the transfer of visual information from one layer to the next, as well as principles followed in the distribution of information, from one cortical area to the next. 3. Descriptions of receptive field properties in every layer at every stage, and insights into transforms that are implied by every change. 4. What these organizations and transforms suggest about the visual processing strategies that are most likely to be pursued.

Gary Blasdel has studied visual neurobiology for the past 25 years. Even though he has investigated a wide range of topics during this time, he has maintained a central interest in visual processing and strategies of feature extraction and scene segmentation that began during discussions of related issues during a physics seminar series at UC Berkeley, and that initiated Dr. Blasdel's interest in vision. He has served on the faculties of four University Medical Schools, and is currently an Associate Professor of Neurobiology at Harvard Medical School. He is best known for a novel imaging technique he developed, that made it possible for the first time to visualize patterns of response preferences in the cortex of living animals.

"Neural Networks for Identification and Control"
Kumpati Narendra, Yale University

Systems theory is a scientific discipline which is inherently interdisciplinary in nature and extends from design, development and production on the one hand to mathematicson the other. Advances in systems theory have been made through a combination of modeling, computation, and experimentation, all of which are essential to the study of neural networks. From a systems theoretic point of view, artificial neural networks can be considered as convenient parametrizations of nonlinear maps from one finite dimensional space to another. Since methods for training such networks using input-output data are currently well known, they are ideally suited to approximate unknown nonlinear functions in differential and difference equations used to rpresent nonlinear dynamical systems. The principal aim of this tutorial is to indicate how results in nonlinear control theory, structures provided by linear adaptive control, and the approximating capabilities of neural networks can be judiciously combined to deal with problems of nonlinear adaptive control that arise in complex dynamical systems.

Kumpati Narendra received the M.S. and Ph.D. degrees from Harvard in 1955 and 1959 respectively. He joined the Department of Engineering and Applied Science at Yale in 1965. He is currently the Director of the Center for Systems Science at Yale. Dr. Narendra is the author of over 150 publications in systems theory, and of three books. His research interests are in stability theory, adaptive control, learning automata, and control of complex systems using neural networks.

Session II: 1:00--3:00 p.m.

"Cortical Circuits in a Multichip Communication Framework"
Misha Mahowald, Institute for Neuroinformatics

The circuits of the neocortex are composed of large numbers of neurons that are highly recurrently connected at both a columnar and areal levels. Incorporating a large number of neurons in analog VLSI requires a number of chips. However,multichip system design for neuromorphic computing is hindered by the problem of communication among chips in the system. A number of research groups have begun to use a digital event based method for interchip communication, such as the Address-Event Representation (AER) or Virtual Wires. In this tutorial, I will present biological data elaborating the characteristics of cortical circuits and demonstrate how such circuits are embedded in a reconfigurable communications framework. I will describe circuits for orientation selectivity and stereo disparity computations and report on progress towards realizing these circuits in a general purpose neuronal emulator.

Misha Mahowald received her PhD. in 1992 from the department of Computation in Neural Systems at Caltech under Carver Mead. She did post-doctoral work with Rodney Douglas and Kevan Martin at the Medical Research Council Anatomical Neuropharmacology Unit in Oxford, England. She is currently at the Institute for Neuroinformatics in Zurich, Switzerland.

"Computational Learning and Statistical Prediction"
Jerome Friedman, Stanford University

A learning system is a computer program that constructs rules for predicting the values of particular aspects of a real system (outputs), given the values of other aspects of that system (inputs). Learning systems attempt to construct useful prediction rules purely by processing data taken from past successfully solved cases; that is, cases for which the values of both the outputs and the inputs have been determined. Methodology and theory for computer learning have been traditionally developed in the fields of applied mathematics (function approximation), statistics (multiple regression and classification), and engineering (pattern recognition). More recently renewed interest and excitement has been generated by research in artificial intelligence (machine learning) and the neurosciences (neural networks). This tutorial will cover the fundamental principles underlying the paradigms of computer learning developed in these fields with the goal of placing them in a common perspective and providing a unifying overview. The presentation will be largely intuitive in nature without recourse to involved mathematics.

Jerome H. Friedman is Professor of Statistics, Stanford University, and the leader of the Computation Research Group at the Stanford Linear Accelerator Center. His research interests have centered on computational learning methodology for over twenty years. He is the inventor or co-inventor of several widely used approaches including the CART, projection pursuit, MARS, and ACE algorithms.

Session III: 3:30--5:30 p.m.