Invited Speakers
Nature's Gift to Science
Sydney Brenner is one of the past centuryís leading pioneers in genetics and molecular biology.† Among his many notable discoveries, Brenner established the existence of messenger RNA and demonstrated how the order of amino acids in proteins is determined. He also conducted pioneering work with the roundworm, a model organism now widely used to study genetics. His research with C. elegans garnered insights into aging, nerve cell function and controlled cell death, or apoptosis. He was awarded the Nobel Prize in 2002.
He is now a Distinguished Professor at The Salk Institute, in La Jolla, California, where he is pioneering a new approach in Computational Biology called the CellMap.
Anders Dale, Harvard University
http://www.nmr.mgh.harvard.edu/NewFiles_Staff/dale_anders.html
Relating Brain Imaging Signals to Biophysical Models of Neuronal Circuits
Abstract:: The goal of the research presented is to integrate information from different imaging modalities in order to obtain estimates of brain activity with optimal spatial and temporal resolution, and ultimately to relate noninvasive imaging signals to biophysical models of neuronal circuits. The problem is phrased in a Bayesian framework, in which three primary forms of information are encoded: 1) the forward models for the different imaging signals, specifying the coupling of the signals with the physiological variables; 2) the coupling between different physiological parameters, such as membrane potentials / synaptic currents and hemodynamics / metabolism; and 3) a priori information about the spatial patterns and dynamics of electrical activity.
High-resolution structural MRI data is used to obtain detailed models of the anatomy of the cortex and other brain structures, providing a priori information about the possible location and orientation of synaptic currents. A combination of multi-spectral structural MRI and Diffusion Tensor Imaging is used to obtain accurate forward models for EEG/MEG and optical imaging signals, and fMRI. Finally, the coupling between local current source density and hemodynamic variables (blood flow, volume, and oxygenation) is encoded in the form of a probabilistic spatiotemporal transfer function estimated from simultaneous electrophysiological and hemodynamic recordings.
Bio: Anders M. Dale received his PhD in 1994 from the University of California, San Diego in Cognitive Science. He is currently an Associate Professor in Radiology at Harvard Medical School /MGH, and is an Associate Director of the MGH/MIT/HMS Athinoula A. Martinos Center for Biomedical Imaging. His main research focus is on the development and utilization of techniques for non-invasive functional and structural imaging of the human brain. A major goal of this research is to develop methods for obtaining high-resolution spatiotemporal estimates of brain activity by combining functional magnetic resonance imaging (fMRI) with magnetoencephalography (MEG) and electroencephalography (EEG) and optical imaging techniques. A long-term goal of this research is to relate non-invasive brain imaging signals to biophysical models of the underlying neuronal circuits.
Michale Fee, Bell Labs, Lucent Technologies
Time and Sequence in the Brain: Insights from a Songbird
Abstract: Sensitivity to temporal order is a nearly universal aspect of brain function - at the sensory, motor, and cognitive levels. The ability of the brain to step rapidly through a learned sequence of states underlies not only the performance of complex motor tasks such as speech, but perhaps our ability to think and plan as well. While little is known about the biophysical and circuit mechanisms underlying the generation and learning of sequences, songbirds provide a marvelous animal model in which to study these phenomena. Using newly developed microscale devices for monitoring the activity of single neurons in the brain of singing bird, we are beginning to understand the circuits that generate complex vocal sequences. Analysis of the dynamics within these circuits show that song is controlled by a group of neurons that form a sparse representation of time in the brain. Such a representation eliminates the problem of interference between different states in the sequence, resulting in faster learning than can be achieved with a more distributed representation.
Bio: Michale Fee did his undergraduate work at the University of Michigan where he studied engineering physics, with a particular interest in lasers and optics. He went to Stanford University for graduate school and worked in the lab of Steven Chu where his thesis work involved precision two-photon measurement of the1S-2S transition energy in positronium. The experiments were done at Bell Laboratories, where Michale got mixed up with a remarkable group of neuroscientists, and was offered a postdoctoral position in the lab of David Kleinfeld. There, Michale helped develop techniques for spike sorting and chronic recording, and studied neurons in whisker barrel cortex of the behaving rat to understand how motor and sensory signals are integrated during active sensation.
Continuing as a principal investigator at Bell Laboratories, Michale became interested in the song control system, which enables songbirds to learn and generate complex vocal motor sequences by listening to a tutor and then practicing their song. Of particular interest is the circuitry that codes and generates temporal sequencing of these vocal gestures. His lab has recently found neurons in the premotor song control circuit that generate only a single brief burst in the sequence, and may form an explicit representation of time in the brain. Michale is also interested in developing new techniques for recording electrical and optical signals in awake behaving animals. Recently developed techniques include a 1.5 gram motorized microdrive for chronic recording, an active electrode stabilizer for intracellular recording in awake animals, and a miniature two-photon microscope for intracellular imaging in freely behaving animals.
Marc Mezard, UniversitÈ de Paris Sud
http://ipnweb.in2p3.fr/~lptms/membres/mezard/
Analytic and Algorithmic Solutions of Random Satisfiability Problems
Abstract: One version of the famous satisfiability problem in computer science asks whether a random Boolean expression built from many clauses with K variables per clause can be satisfied by an appropriate assignment of the variables. This NP-complete problem is well known to be computationally hard near to the threshold value of the ratio of clauses-to-variables separating cases which are generically satisfiable (below the threshold), from those which are generically unsatisfiable (above the threshold). In fact, this algorithmic difficulty is due to the existence of an intermediate glass phase just below the threshold, where metastable states proliferate.
The understanding of this structure has led recently to the introduction of a new type of message passing algorithm, the survey propagation algorithm, which is a kind of generalization of belief propagation, able to deal with the multiplicity of metastable states: it is able to solve large instances of K-satisfiability with millions of variables quite close to the threshold. Taking the satisfiability problem as a guide, the talk will review some important issues at the interface between statistical physics and computer science, focusing on phase transitions and message passing procedures.
Bio: Marc Mezard is a research director at CNRS, working at the "Laboratoire de Physique Theorique et Modeles Statistiques" at the "Universite de Paris Sud". He received a PhD in Theoretical Physics at the Ecole Normale Superieure (Paris) where he was hired after a post-doctoral stay at Rome University. In 2001 he joind the Universite de Paris Sud. The stem of his research is the statistical physics of disordered systems; he has been working on various aspects of glassy phases in physical systems, and on emergent properties in large systems of interacting heterogeneous "atoms", with applications to neural networks, combinatorial optimization problems and financial markets.
Elissa Newport, University of Rochester
http://www.bcs.rochester.edu/people/newport/newport.html
Statistical Language Learning in Human Infants and Adults
Abstract: In collaboration with Richard Aslin, I have recently been developing an approach to language acquisition known as 'statistical learning' (Newport & Aslin, 2000, and in progress). The central notions are a blend of ideas from structural linguistics (Harris, 1951) and nativist perspectives (Chomsky, 1955, 1981) with recent proposals using distributional analysis in language acquisition. Our basic proposal is that important parts of language acquisition may involve learners computing, over a corpus of speech, such things as how frequently sounds co-occur; how frequently words occur in similar contexts; and the like. These computations may then be used to determine regular versus accidental properties of the language being acquired. Our studies (initially in collaboration with Jenny Saffran) have shown that human adults and infants are capable of performing many of these computations online and with remarkable speed, during the presentation of controlled speech streams in the laboratory. We have also found that adults and infants can perform similar computations on nonlinguistic materials (e.g., music), and (in collaboration with Marc Hauser) that nonhuman primates can perform the simplest of these computations. However, when tested on more complex computations involving non-adjacent sounds, humans show strong selectivities (they can perform certain computations, but fail at others), corresponding to the patterns which natural languages do and do not exhibit. Primates are not capable of performing some of these more difficult computations. This approach may provide an important mechanism for learning certain aspects of language. In addition, the constraints of learners in performing differing types and complexities of computations may provide part of the explanation for which learners can acquire human languages, and why languages have some of the properties they have.
Bio: Dr. Elissa L. Newport is the department chair and the George Eastman Professor of Brain and Cognitive Sciences at the University of Rochester. Her primary research interest is in human language acquisition, with research projects including naturalistic studies of children learning their first languages, experimental studies of infants, adults, and non-human primates learning miniature languages in the lab, fieldwork on emerging sign languages, and fMRI research on language and the brain. Professor Newport received her PhD in Psychology at the University of Pennsylvania and was a Sloan Fellow in Linguistics and Cognitive Science at Penn and MIT. She has been on the faculty at the University of California at San Diego, University of Illinois, and, since 1988, University of Rochester. Her research is funded by the NIH, NSF, the McDonnell Foundation, and the Packard Foundation, and for this research has received the Claude Pepper Award of Excellence from NIH. She currently is a series editor for MIT Press and serves on the Board on Behavioral, Cognitive, and Sensory Sciences of the National Academy of Sciences, and she is a Fellow of the American Academy of Arts and Sciences and the American Association for the Advancement of Science.
David Salesin, University of Washington and Microsoft Research
http://salesin.cs.washington.edu/
The Need for Machine Learning in Computer Graphics
Abstract: Machine learning has the potential to revolutionize the field of computer graphics. Until now, so many of the successes in computer graphics from realistic plant models to human animation to cinematographic effects have been achieved, painstakingly, through the creation of highly complex models by hand. Unfortunately, the process for creating these models does not scale: Whether for plants, animation, or cinematography, good models are hard to come by, with each model having to be crafted, individually, by an expert. Yet good examples of all of these things are all around us. Thus, for computer graphics to achieve its full potential, what we really need is for all of these highly complex models to be constructed automatically from the examples themselves in other words, what we really need is machine learning! In this talk, I will survey some early applications of machine learning to computer graphics and present dozens of new challenges for future research.
Bio: David Salesin is a Professor in the Department of Computer Science and Engineering at the University of Washington, where he has been on the faculty since 1992, and a Senior Researcher at Microsoft Research, where he has also worked since 1999. He received his ScB from Brown University in 1983, and his PhD from Stanford University in 1991. From 1983-87, he worked at Lucasfilm and Pixar, where he contributed computer animation for the Academy Award-winning short film, ``Tin Toy,'' and the feature-length film ìYoung Sherlock Holmesî. During his years at Stanford, he also worked as an intern at the DEC Systems Research Center and Paris Research Lab. He spent the 1991-92 year as a Visiting Assistant Professor in the Program of Computer Graphics at Cornell University. He has consulted at Sogitec Audiovisuel, Aldus (now part of Adobe), Xerox PARC, Broderbund, and Microsoft Research. In 1996, he co-founded two companies, where he served as Chief Scientist: Inklination and Numinous Technologies (acquired by Microsoft in 1999).
Salesin received an NSF Young Investigator award in 1993; an ONR Young Investigator Award, Alfred P. Sloan Research Fellowship, and an NSF Presidential Faculty Fellow Award in 1995; the University of Washington Award for Outstanding Faculty Achievement in the College of Engineering in 1996; the University of Washington Distinguished Teaching Award in 1997; The Carnegie Foundation for the Advancement of Teaching and the Council for the Advancement and Support of Education Washington Professor of the Year Award in 1998; the ACM SIGGRAPH Computer Graphics Achievement Award in 2000; and he became an ACM Fellow in 2002. Salesin's research interests are in computer graphics, and include, in particular, non-photorealistic rendering, image-based rendering, and various topics in 2D graphics like color reproduction, digital typography, and compositing.