a.jennings at trl.OZ.AU
Thu Feb 25 15:53:59 EST 1993
I'm very interested in your comments. It certainly doesn't seem sensible to
download things when there is a large variety of response.
In a system like the brain where we want to trade storage for speed, isn't it
better to send intermediate results off to the parts of the brain that can
best interpret them?
I don't think "downloading" is good for wide-variety response functions.
Which leaves us the question of how its <really> done.
HoHum. Its really difficult.
In the brain, the need is for rapid response to the sensory environment.
To optimize that, the senses have to report on objects in the animal's
internal model as directly as possible. Otherwise there's a lot of
correlation that has to be done for each breath, each eye staccade, and
each sound. I speculate that "downloading" is performed in the olfactory
system by the AON presenting synthetic sensory data to the "wet end" of
the OB while pulling the back end of the OB to the hyperbolic point
that the PPC uses as the starting point for processing when a breath
is taken. The output of the entire system is a "noisy limit cycle"
(NLC) that identifies the object being detected. This NLC consists
of an underlying carrier wave modulated by the novel data associated
with the object. This combined signal (which is evidenced to the cortex
by a pattern of neurons that have become synchronized and periodic
in their spike trains) is generated by either the OB or by the PPC.
If generated by the OB, the AON or the PPC has to simultaneously present
the synthetic sensory data and the carrier wave during training. If
generated by the PPC, then the PPC has to learn to correlate the output
of the OB generated during training with the carrier wave to be used for
There are a number of ways to combine the carrier wave with the object
data. One speculation is that the carrier wave is continuous, while the
object data consists of bursts marking the components of the carrier wave
where there was matching. It could be the other way around, too.
Why not download a large number of objects to the OB permanently? As an
engineer who builds similar systems, I suspect there are resource
management problems. If too many objects are downloaded, convergence might
be delayed or made impossible. There may also be processor limitations in
the OB/PPC system. Or there may be far too many -->possible<-- objects
that could be downloaded to the OB/PPC.
Note that one implication is that semantic objects are represented by
synchronized periodic patterns of excitation.
Internet: erwin at trwacs.fp.trw.com
More information about the Neur-sci