rmyers@ics.uci.edu (Richard E. Myers) (10/11/90)
I am looking for information on simulation packages that can be used
to model neural networks from the bottom up and possibly using
anatomical data. We now use a modified version of the Rochester
Connectionist Simulator (RCS). An ideal replacement would have the
following attributes:
OBJECT ORIENTED - Models should be implemented using an object
oriented paradigm. For example, a model of a "brain" would be
composed of many "neuron" objects that each contain their own
state and connectivity data but share the same functionality.
This would help in building bottom up models because low level
objects such as "neurons" could be debugged separately before
you assemble a "brain".
EVENT DRIVEN - The sequence of events in the simulator should
be dynamically generated and scheduled by the objects being
modeled. For example, object B causes object C to change only
if object A causes a change in object B that exceedes a
certain threshold. The alternative would be to pass control to
all three objects in order and have each object check if it
should change based on the results of its predecessor.
PSEUDO TIME - Objects should be able to schedule the order of
events with some notion of real time. Imagine the previous
example but with a "delay" before control passes to object C.
Control would pass to other objects before object C changes.
I believe this is a common feature of data network simulators
in which the delay caused by a signal propagating down a wire
is important.
We would also like a system with a graphical user interface that
allows for easy monitoring of events and probing of objects. A
revision control system and an experiment control system would also be
nice. :-)
POSSIBLE SOLUTIONS
GENESIS comes very close to what we want but it lacks a true
(small-talk) oriented approach to creating models. Communication
between objects also appears pretty limited and predefined.
BONES might work but I haven't heard of anyone successfully modeling
neural networks with it. I have heard that BONES has a nice revision
control and experiment control system though.
RCS isn't object oriented, event driven or have any concept of time.
I welcome any changes or extensions to the above specification or
any comments or alternatives to my analysis of possible solutions. I
realize that I'm asking for the moon, but hopefully this will generate
some discussion on what is available and what is needed for neural
network simulation.
Thank you for your time,
-- Richard
-------------------------------------------------------------------------------
"Programs were devised whose `thinking' was to human thinking as a
slinky flipping end over end down a staircase is to human locomotion."
-- D.R. Hoffstadter
-------------------------------------------------------------------------------
Richard Myers / 6 Exeter / Irvine, CA 92715 / 714-854-4410 / rmyers@ics.uci.edu
--
-------------------------------------------------------------------------------
"Programs were devised whose `thinking' was to human thinking as a
slinky flipping end over end down a staircase is to human locomotion."
-- D.R. Hoffstadter