*********************************************************************
*                SEMINAR  "NEUROBIOLOGIE UND BIOPHYSIK"             *
*********************************************************************

                  DONNERSTAG, 14. Mai 1998

                                   17:15 Uhr
                    Hansa-Str. 9a, Hoersaal EG

                       Prof.  Charles H. Anderson
                 Dept. Anatomy and Neurobiology
          Washington University School of Medicine
                       St. Louis, MO 63124, USA

    "UNIFYING PERSPECTIVES ON NEURONAL POPULATION CODES"

Abstract:
A principled approach to neuronal population codes is developed as a
step toward providing a general computational framework for
understanding neurobiological systems. The discussion begins with the
Georgopoulos Population Vector, a linear decoder where the value of an
encoded vector is estimated using a weighted sum of the neuronal
activities. The optimal values for the prefered directions are found
using a least squares estimator (LSE) [Salinas and Abbott 1994].
Bialek [1997] and colleagues have also found that a linear decoder can
estimate a temporal signal encoded by a spike train, where the
decoding filter is found using a similar LSE. The simplest
population-temporal code (PT) is obtained by combining these two
linear decoders. Indeed, the analyses of temporal decoding using the
fly's H1 neuron [Bialek] and the cricket cercal system [J. Miller
1991] are based on a population of 2 (`on/off') neurons.  This linear
PT code can be generalized to represent densities functions that allow
for encoding multivalued estimates of parameters such as stereo depth.
Encoding and decoding functional pairs based on a LSE allow densities
to be encoded into a PT code of spike trains that can be decoded with
a precision that increases with the number of neurons. This analysis
of precision of representation leads to a generalized class of
attractor networks [S. Seung 1996 and K. Zhang 1998], which can be
used as a form of working memory for analog quantities.  Computational
networks of spiking neurons can then be explicitly constructed from
weighted averages of conditional probabilities that appear to have
many properties of real neuronal circuits. This framework is also
consistent with many recent developments in the application of
Bayesian techniques to complex problems in statistical inference, such
a Pearl's Belief Nets and Grenander's Pattern Theory [M. I. Jordan
1996].

Some References:

C. H. Anderson and D. C. Van Essen, "Neurobiological Computational
Systems", In: Computational Intelligence, Imitating Life, Ed. J. M
Zurada, R. J. Marks II, C. H. Robinson, IEEE Press, pp 213-222, 1994.

Salina and Abbott, "Vector Reconstruction from Firing Rates",
J. Computational Neuro. 1, 89-107 1994.

R. Rieke, D. Warland, R. R. van Steveninck, W. Bialek, "Spikes,
exploring the neural code", MIT Press, 1997.

J. P. Miller, G. A. Jacobs, and F. Theunissen, "Representation
of Sensory Information in the Cricket Cercal Sensory System",
J. Neurophysiology, 66, 1680-1703, 1991.

S. Seung, "How the brain keeps the eyes still", Proc Natl.Acad. Sci.
USA, 93, pp13339-13344, 1996.

K. Zhang, I. Ginzburg, B. L. McNaughton, and T. J. Sejnowski,
"Interpreting Neuronal Population Activity by Reconstruction:",
The American Physiological Society, pp 1017-1044, 1998.

M.I. Jordan and Christopher M. Bishop, "Neural Networks",
A.I. Memo No. 1562, MIT Artificial Intelligence Laboratory, March
1996.

A. Pouget, K. Zhang, S. Deneve, and P. E. Latham, "Statistically
Efficient estimation using population codes", Neural Comp. 10, pp
373-402, 1998.

R. S. Zemel, P. Dayan, and A. Pouget, "Probabilistic interpretation of
population codes", Neural Comp. 10, pp 403-430, 1998.

Note: Lecture announcements including abstract can also be found at our
website: http://www.brainworks.uni-freiburg.de/calendar/forthcoming.html