Large scale biologically realistic simulations

[Dan Hammerstrom]

Dear Dr. Hammarlund:

We noticed your recent posting on the Connectionists' network mailing
list.  We feel we may have some data that would interest you.

We are currently developing a neurocomputer system designed for the
high computational demands of neural network simulations.  Our system
is based around custom VLSI circuits organized in a SIMD (Single
Instruction, Multiple Data) architecture.  A single board has a peak
capability of 12.8 billion multiply accumulate operations per second.

In addition to developing a variety of neural network algorithms for
this machine, we are also developing, under contract from the Office
of Naval Research and in conjunction with the Oregon Graduate
Institute and the Center for Learning and Memory at the University of
California at Irvine, a real time simulation of a 10,000 neuron slice
of olfactory pyriform cortex.

The model we are using is that of Gary Lynch and Rick Granger and
thier colleagues at UCI.  Although it is much more abstract, and less
realistic than the models you are working on, it still has a number of
interesting properties including the ability to form hierarchical
categorizations of the input vector space in an unsupervised mode.
 
In our implementation, a linear array of piriform layer II neurons is
mapped to a linear of array of processors.  Up to 512 neurons can
execute simultaneously in the neurocomputer; this ``slice'' of the
network would constitute a single stage of a processor pipeline.
Selectively piping the outputs of one slice as inputs to the next
would emulate the feed--forward character of the Lateral Olfactory
Tract, or LOT.  We estimate that a piriform network consisting of
approximately 10,000 pyramidal cells and a 512--element LOT could
process 500 million eight--bit connections per second.  At 10% sparse
LOT connectivity, the entire system has roughly 750,000 synapses when
lateral inhibitory connections are included.  This performance is
equivalent to roughly 1000 LOT presentations per second, or 200
presentations per second when the network is in learning mode.  (Which
is significantly faster than real time, and allows for some
interesting research.)

Researchers at the Oregon Graduate Institute are currently
investigating the applications of the piriform model executing on such
speech--processing applications.
 
Please refer to the following for specific information concerning our network 
architecture and the piriform model (forgive the Latex format):

@ARTICLE{piriform,
        AUTHOR = "Gary Lynch and Richard Granger and J{\'o}se Ambros-Ingerson",        TITLE = "Derivation of encoding characteristics of layer 2 cerebral
        cortex",
        JOURNAL = "Journal of Cognitive Neuroscience",
        YEAR = {1989},
        VOLUME = {1},
        NUMBER = {1},
        PAGES = {61--87}  }

@INCOLLECTION{memorial,
        AUTHOR = "Richard Granger and J{\'o}se Ambros--Ingerson and Ursula
        Staubli and Gary Lynch",
        TITLE = "Memorial Operation of Multiple, Interacting Simulated Brain
        Structures",
        BOOKTITLE = "Neuroscience and Connectionist Models",
        PUBLISHER = {Erlbaum Associates},
        YEAR = {1989},
        EDITOR = "M. Gluck and D. Rumelhart" }
 
@INPROCEEDINGS{AdaptiveIJCNN,
        AUTHOR = "Dan Hammerstrom",
        TITLE = "A VLSI architecture for high-performance, low-cost,
        on-chip learning",
        BOOKTITLE = "Proceedings of the 1990 International Joint Conference
        on Neural Networks",
        YEAR = {1990},
        MONTH = {June}  }
 

Sincerely,

Eric Means, Adaptive Solutions
Dan Hammerstrom, Adaptive Solutions / Oregon Graduate Institute
Todd Leen, Oregon Graduate Institute