[Fwd: Re: [Comp-neuro] Noise, redundancy and biophysics]

Ali Minai minai_ali at yahoo.com
Thu Jul 24 18:00:49 CEST 2008


Jim, your point is well-taken, but in fairness, the question to ask is whether that original - perhaps somewhat misguided - infatuation with the Hopfield network hindered or promoted work on more complex models, leading to the more realistic ones we see today. I think that it helped by bringing in a whole community of very smart people - physicists, computer scientists, mathematicians, engineers - some of whom then took the time to learn the biology and eventually made great contributions closer to the level you like. True, they also created a lot of noise, but haven't we all been agreeing that noise is good?:-)

I know, of course, that physicists and mathematicians did neuroscience long before Hopfield networks, but not in the numbers they have since. On balance, I think that elegant if not directly applicable ideas such as Hopfield networks, supervised learning, clustering, ICA, etc., have greatly enriched neuroscience by seeding ideas.

Cheers

Ali


---------------------------------------------------------------------
Ali A. Minai
Complex Adaptive Systems Lab
Associate Professor
Department of Electrical & Computer Engineering
University of Cincinnati
Cincinnati, OH 45221-0030

Phone: (513) 556-4783
Fax: (513) 556-7326
Email: aminai at ece.uc.edu
          minai_ali at yahoo.com

WWW: http://www.ece.uc.edu/~aminai/

----------------------------------------------------------------------

--- On Thu, 7/24/08, jim bower <reinoud at castafiore.cde.ua.ac.be> wrote:
From: jim bower <reinoud at castafiore.cde.ua.ac.be>
Subject: [Fwd: Re: [Comp-neuro] Noise, redundancy and biophysics]
To: comp-neuro at neuroinf.org
Date: Thursday, July 24, 2008, 5:07 AM

---------------------------- Original Message ----------------------------
Subject: Re: [Comp-neuro] Noise, redundancy and biophysics
From:    "jim bower" <bower at uthscsa.edu>
Date:    Wed, July 23, 2008 7:49 pm
To:      "Sacha Nelson" <nelson at brandeis.edu>
         comp-neuro at neuroinf.org
--------------------------------------------------------------------------

Not clear that the climbing fiber has any real significance for the output
of the cell. For its neutron bomb effect on the dendrite, it produces 1 or
2 spikes as output.

My own view is that it is mostly a dendritic phenominon and also has
nothing to do with learning.

Yes, small cells with no dendrites do something different than large cells
with big dendrites,  but even the neurons in nucleus lamineris are not
well understood.

Granule cells have small dendrites, but they stick into a glomerulous
which is likely enormously complex.

So, generally speaking, generic neuronal processing isn't likely to be
found anywhere when you look closely.

I am reminded of the first neural network meeting I attended in 1984 at
the miramar hotel in Santa Barbara. In fact it was the second meeting of
the modern era. And the room was abuzz with talk of the  hopfield network
and traveling salesman (and spin glasses). I was asked numerous time if
there were Hopfield networks in the brain. What I said was no, and
therefore it was very likely that artificial neural networks would have to
get much more complex and messy to do anything practically interesting.
This opinion was definately not well received as the elegance of the
Hopfield network, its energy functional, and link to well known theory
were seductive properties, and physicists hate complexity. ---  25 years
later. - you be the judge.

;-)

Jim
Sent via BlackBerry by AT&T

-----Original Message-----
From: Sacha Nelson <nelson at brandeis.edu>

Date: Wed, 23 Jul 2008 10:54:37
To: <bower at uthscsa.edu>
Subject: Re: [Comp-neuro] Noise, redundancy and biophysics



> Sacha,
>
> I agree thus it would be nice to agree on a common definition -
> which has been very hard to do.
>
> A couple of questions:
>
> How can we possibly know what is "relevant" to a particular
neuron?
> We can decide based on our experimental protocohls, but isn't that
> us imposing on "them".

I think it is not so much relevant to a particular neuron as relevant
to a particular computation. The ultimate arbiter of relevance must
always be behavior. Experimentally, it is incredibly hard in most
cases to make this causal link: this signal transmitted from this
neuron to this other neuron is necessary for the execution of this
action or perception, but conceptually I think the acid test is
straightforward. I believe that almost any signal that a
neurophysiologist could correlate in my brain with some aspect of a
stimulus, I could learn to use to discriminate those stimuli and
modify my behavior accordingly. But from the point of view of most of
my behaviors those signals might be "noise."

>
>
> "the reliability and temporal precision
> with which a single presynaptic action potential results in a
> postsynaptic action potential can be low". In the cerebellum, the
> 150,000 excitatory inputs to each Purkinje cell don't appear to have
> any direct influence on the output of the cell - I suspect most
> excitatory synaptic inputs in the brain are actually doing exactly
> what they appear to be doing, nfluencing the local membrane. I
> believe that we neurobiologists may be far too "soma-centric" in
> thinking about how neuons and brains compute.

yes, but purkinje cells and the apical tufts of pyramidal neurons are
somewhat specialized cases. At the other end of the extreme are bushy
cells or granule cells. Individual granule cells may not be doing much
from the perspective of somatic voltage, but the climbing fiber input
certainly is.


>
>
> Last isn't it intersting that the closer one gets to the periphery
> either on the sensory or the motor side the more "precise" the
> nervous system looks, yet somehow in the middle it looks like it
> needs to solve some signal to noise problem. But why isn't it simply
> loikely that we can better understand what is going on on either
> end, but in the middle we have no idea.

I agree it is harder to figure out what is going on in the middle, but
it is really about bottlenecks and the relative importance of pure
transmission vs. more complicated computation. The retinal ganglion
cell axon is actually many synapses away from the photoreceptor, but
it is specialized for transmission. Closer to the periphery, exactly
what is going on in some amacrine and horizontal cell networks has
been incredibly difficult to figure out. The job of the ganglion cell
axon (like the climbing fiber) is essentially very different from a
recurrent excitatory synapse in the cortex or a granule cell input to
a purkinje cell. The experiment of recording from a pre and
postsynaptic cell and asking about the reliability, amplitude etc. of
transmission can be done anywhere. Ascribing sensory or motor meaning
to these signals is of course more difficult, but that is a different
business.

>
>
> Although I know bard and others who talk about noise are talking
> about it in a precise way -- however, I tend to lump the word
> "noise" like a number of other similar words in literature as a
bin
> in which we put things we don't understand.
>
> I am reminded of Penzias (sorry about the spelling and wilson
> crawling around in their satalite dish with tin foil trying to
> remove what turned out to be (predicted) background cosmic radiation.
>
> (Please don't point out the role of theoretical modeling in
> realizing the truth. -- this modeling was done in the context of a
> science of simple things (physics) with among other things a common
> set of definitions).
>
> Jim
> Sent via BlackBerry by AT&T
>


_______________________________________________
Comp-neuro mailing list
Comp-neuro at neuroinf.org
http://www.neuroinf.org/mailman/listinfo/comp-neuro
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://www.neuroinf.org/pipermail/comp-neuro/attachments/20080724/6f97ad4d/attachment-0001.html


More information about the Comp-neuro mailing list