Boulder Computational Learning Group
Fall 1997 Schedule
Unless otherwise noted, meeting time is on Wednesday, 3-5 p.m.,
in ECOT 831a (Engineering center, 8th floor, small conference
room). Dates in red are at the usual time and place.
Dates in blue are special meeting dates.
September 4, 7:30 p.m., Old Main Chapel: Daniel Schacter, Harvard
The cognitive neuroscience of illusory memories
Memory illusions and distortions have long been of interest to
psychological students of memory, but neuropsychologists and
neuroscientists have paid relatively little attention to them. This talk
will present a cognitive neuroscience analysis of memory illusions and
distortions by considering evidence from a patient with a right frontal
lobe lesion, patients with amnesia produced by damage to the medial
temporal lobes, normal aging, and healthy young volunteers studied with
functional neuro-imaging techniques. Particular attention is paid to the
contrasting roles of prefrontal cortex and medial temporal lobe structures
in accurate and illusory remembering, and to the mechanisms that are
involved in reducing or suppressing powerful but false recollections.
September 17
In anticipation of Geoff Hinton's visit, we will read and discuss
several of his recent papers that will be relevant to his talks.
In order of discussion:
September 18, 4:00 p.m., Muenzinger D430: Shaun Vecera, University of Utah
September 19, 3:30 p.m., Muenzinger D430: Bernard Baars, Wright Institute
Neural hypotheses derived from Global Workspace Theory:
Elements of a cognitive neuroscience of consciousness
Global Workspace theory is a simple cognitive architecture that has been
developed to account qualitatively for a large set of matched pairs of
conscious and unconscious processes (Baars, 1983, 1988, 1993, 1997). Such
matched contrastive pairs of phenomena can be either psychological or neural.
Psychological phenomena include subliminal priming, automaticity with
practice, and selective attention. Neural examples include coma and
blindsight. Like other cognitive architectures, GW theory may be seen in
terms of a theater metaphor of mental functioning. Consciousness resembles a
bright spot on the theater stage of Working Memory (WM), directed there by a
spotlight of attention, under executive guidance. The rest of the theater is
dark and unconscious. All the elements of GW theory have reasonable brain
interpretations, allowing us to generate a set of specific, testable brain
hypotheses about consciousness and its many roles in the brain. This
approach is compatible with a number of other proposals.
September 24
Continue discussion of Hinton papers
September 25, 3:30 p.m., ECCR 265: Geoff Hinton, University of Toronto
Fitting intractable generative models
When fitting a stochastic generative model to data it is normal to compute
the posterior probability distribution that is induced over configurations
of the hidden variables by each observation. For simple models such as
mixtures of Gaussians and factor analysis this posterior distibution can be
computed exactly. For more interesting generative models composed of
multiple layers of non-linear units it is intractable to compute the
posterior distribution. I shall describe various ways of approximating the
posterior distribution and show that simple learning rules can improve the
generative model even when the approximations are poor.
September 26, 3:30 p.m., Muenzinger D430: Geoff Hinton, University of Toronto
A hierarchical community of experts
Perceptual systems learn to convert unlabeled sensory data into explicit
representations of the true causes of the data. This type of unsupervised
learning can be performed by fitting a hierarchical generative model which
captures redundancies in the data at many levels. I shall describe a novel
type of multilayer generative model which consists of many linear experts
and a non-linear mechanism for selecting an appropriate subset of the
experts to model each observation. When trained on images of handwritten
digits the model learns interesting and appropriate features and discovers
digit classes without supervision. This is joint work with Brian Sallans
and Zoubin Ghahramani.
October 1: NO MEETING
October 8: Mike Mozer
Modeling subliminal perception
October 9, 5 p.m., Room 158, Department of Psychology,
University of Denver (2155 S. Race St.): Randy O'Reilly
Biological Principles for Simulating Cognition: Learning & Memory
October 15: Dietmar Ruwisch
Visiting Heinz Willebrand (heinzw@spot) 10/5-10/18
Wave propagation as a neural coupling mechanism: Hardware for self-organizing
feature maps and the representation of temporal sequences
Wave propagation within a network of neurons can be used to
control the learning process of the neurons in a way, that the network
generates self-organizing features maps (Kohonen network).
Furthermore, the propagating wave can be used to influence the neural
competition in representing the input of the network. By this means the network
is able to represent temporal aspects of the input.
Since apart from a global bus the network only requires local interactions,
connectivity is very low. Thus, the concept is well-suited for a
parallel hardware implementation. Operation of a digital demonstration setup
consisting of 16 neurons is demonstrated by the representation of phoneme
sequences.
October 22: Randy O'Reilly
Current research activities
October 29: Soo Young Lee
Current research activities
October 30, 3:30 p.m., ECCR 265: Michael Kearns, AT&T
Quantifying Induction: Some Recent Developments in the Theory of Machine Learning
The exchange of ideas between theoreticians and practitioners of
machine learning has risen dramatically in the last decade. In this
talk, I will sketch developments in two of the topics in which this
cross-fertilization has been the most effective and interesting: the
study of learning curves, which quantifies the rate at which a learning
algorithm generalizes from empirical data, and the framework known as
boosting, which has led to the discovery of new algorithms and shed
light on some classical heuristics. Along the way, I will highlight
some of the many areas that have had direct and powerful impact on
modern machine learning, including information theory, statistical
mechanics, and combinatorics.
November 5: Bernd Olleck and Marc Pickett
Current research activities -- reinforcement learning
November 13, 5 p.m., Room 158, Department of Psychology,
University of Denver (2155 S. Race St.): Linda Barlow, DU
A Taste For Nervous System Development
November 19: Michael Korkin, Genobyte
Genobyte, Inc. is developing a large-scale evolvable hardware system for
ATR in Kyoto. This system will evolve three-dimensional neural networks based
on cellular automata, and run an artificial brain consisting of 1-2 million
neurons in real time to control a robotic device. More information is
provided on our web site www.genobyte.com (under "related publications"), and
in the September 26, 1997 issue of Science Magazine.
December 8, 9 a.m., Room 158, Department of Psychology,
University of Denver (2155 S. Race St.): Daniel Bullock, Boston University
Neuronal and Neural Network Dynamics in the Composition of Action
Many casual observers of action suppose that stored motor programs account for
much of our behavior. However, just as psycholinguistic data force the
conclusion that most linguistic utterances are novel constructions generated
on the fly, a close examination of more mundane action, and of data on its
neural bases, forces the conclusion that even the simplest voluntary actions
are composed as they are being produced. In particular, many brain areas that
contribute to action and its perceptual guidance remain active throughout the
interval of movement production. This fact falsifies strictly serial models,
and complicates analysis of the distinct contribution made by each area,
because in many tasks neural activities in the contributing areas look
strikingly similar. This has led to an anomalous situation in which single
cell neurophysiology has recently emphasized population averaging methods,
which can result in a significant loss of information. Clarifying the
situation requires theories of how movement could be composed on the fly with
neural networks conforming to known anatomy, because such theories allow the
prediction of subtle qualitative differences in the activation profiles of
contributing cell types during movement. In this lecture, I will summarize
recent developments in our long-term program to develop such a theory.
December 11, 5 p.m., Room 158, Department of Psychology,
University of Denver (2155 S. Race St.): Yuko Munakata, DU
Rethinking Infant Knowledge: Insights From Connectionist Modeling
March 12: Yoram Singer, AT&T (tentative)
March 12, 5 p.m., Room 158, Department of Psychology,
University of Denver (2155 S. Race St.): Marie Banich, U. Illinois
The role of the corpus collosum in modulating attention