[Beowulf] [neuro] [tt] kurzweilai: human-scale neural network simulation

Alan Louis Scheinine alscheinine at tuffmail.us
Tue Nov 20 09:08:31 PST 2012


I do not work on AI but would like to keep up with recent events.
There was an article in Communications ACM (if I'm not mistaken)
within the last year or two about the brain mapping, and this
most recent SC12/RJ10502.pdf describes the hardware.  I've been
disappointed that there were no links to fundamental insights
gained with this or with smaller prototypes.

In the early 1970's for an undergraduate class paper I wrote
that the cerebellum was similar to a hologram, thereby implementing
an associative memory.  (Very high connectivity 10^6 in, each output
has 10^4 to 10^5 connection, many sources of input both physical and
from cerebrum.  No need to tailor which input connects to which output,
just as no need to control phase of wave from each source of light
to each silver halide grain, with enough connectivity one gets
associative memory.  Maybe my explanation is too brief to be clear.)

Later I learned about "spin glass" and "energy minimization" type of
neural network -- based on the terminology it is obvious that the
course was taught from the perspective of a physicist.  I realized
that the associative memory with feedback gave a general purpose
computer.  Having such a system partition itself into functional
units would result in an awesomely complex and potentially powerful
AI computer.

The above is old news, old ideas.  The question is: for the layered
structure of the cerebrum, what can it do?  Perhaps there are new
paradigms, new broad, conceptual models of brain function.  My point
is, why do descriptions of the hardware not give links to research
groups that are thinking creatively about what the hardware can do.
It is reasonable to suppose that with the capability of simulating
530 billion neurons and 100 trillion synapses and also having a
logical organization similar (in some respects) to a real brain,
there are research groups of really clever people who have written
new paradigms about neural computing and are been doing simulations
on machines that are smaller than the one described in the SC12
article.

Where can I and others learn about the conceptual developments
related to these large neural simulators?

Alan

-- 

  Alan Scheinine
  200 Georgann Dr., Apt. E6
  Vicksburg, MS  39180

  Email: alscheinine at tuffmail.us
  Mobile phone: 225 288 4176

  http://www.flickr.com/photos/ascheinine


More information about the Beowulf mailing list