[Comp-neuro] cognitive representations & abstract modelling

Ross Gayler r.gayler at gmail.com
Wed Jul 30 06:35:33 CEST 2008


Jim,

(I've changed the subject line to reflect the direction of this post.)

> thanks for support from an unexpected source

I'm always happy to help someone out of the black pit of grubby
physiological detail and into the pure, clear light of abstract models. ;-)
But, seriously though ...

> if you are trying to sort these things out in the context of the visual
system, as most are, you may be starting at the wrong end. It seems to me
likely that it is the olfactory system that invented basic cortical
circuitry.

Lucky for me my models are so abstracted that I don't make any distinction
between sensory modalities :-)  My focus is on specific representational and
computational issues rather than physiological implementation.  All I care
about in that direction is that implementation of the models as real systems
should not be blatantly infeasible.  

One of the interesting aspects of the cognitive problem domain is the
essentially open-ended nature of the conceptual space.  New concepts are
being created and used all the time.  Contrast this with a low level sensory
space, say - the retina, where the representation space is fixed by the set
of photoreceptors.  By analogy, running a visual system like a cognitive
sytem would require continually adding new photoreceptors in new locations
with new spectral sensitivities.  The open-endedness of conceptual space
strongly suggests (at least, to me) that the cognitive level of
representation is quasi-independent of the physiological implementation.
That is, the physiological circuitry implements a virtual machine on which
the cognitive processes are executed while remaining relatively independent
of what is happening at the physiological level.  To give a more concrete
example: in Vector Symbolic Architectures you could implement a virtual
"grandmother cell" whose implementation is distributed over a large number
of actual neurons, none of which is dedicated to only that task.  You could
also dynamically create virtual cells for "proton spin", "online social
networking" and "incestuous love triangle as punishment for peeving the
gods" simultaneously implemented over the same actual neurons.  (There is a
little more about this idea at http://cogprints.org/500/)

However, this distributed virtual machine view does not necessarily commit
me to "arguing a la Lashley that there is no structure in cortex and
everything is completely distributed".  I am very partial to Lawrence
Barsalou's "Perceptual Symbol Systems" hypothesis
(http://www.psychology.emory.edu/cognition/barsalou/onlinepapers.html).  He
argues that cognition consists of modal perceptual systems being run as
simulators.  This would be consistent with different cortical areas being
modality specialists (with the possibility of modality-specific
architectures) provided that there is some communication mechanism to allow
integration of the cognitive simulations of the different cortical areas.
VSA provides some room for optimism here.  It deals only with vectors of
activity and is blind to the architecture that generated the activities - so
effective communication between cortical areas with different architectures
would not be out of the question.  However, there is the problem of how one
cortical area learns what the input from another area means.  Without
wishing to appear glib, this is the same problem that a cortical area faces
with respect to its own sensory input.  Every input is just a vector of
activities, so if the architecture can solve the "what does it mean" problem
for one input it should be able to do the same for all inputs.

In a similar vein, comp-neuro readers may be interested in Rich Sutton's
"Predictive State Representations"
(http://www.cs.ualberta.ca/~sutton/Talks/Talks.html#Predictive_Representatio
ns_of_State_and).  PSRs represent the state of an agent's knowledge about
its current situation as a set of predicted future sensorimotor interactions
with the world.  This is a radical claim, as it asserts that *all* knowledge
is repesented in sensorimotor terms (an idea that has been around for some
time, e.g. see http://cogweb.ucla.edu/Abstracts/Johnson_87.html).  PSRs
ought to be appealing to neurophysiologists because they are very concretely
tied to sensorimotor events.  At the same time, they cry out for a
virtualising implementation (like VSA) because they are predictions of
sequences of events that have not yet happened and because the agent must be
able to represent many such sequences simultaneously on the same
physiological implementation.

So, returning to your initial statement that "it is quite likely that
individual neurons in at least the mammalian brain don't matter" - I agree
on the grounds that it appears to be a requirement flowing from having a
cognitive system.

Ross Gayler



-----Original Message-----
From: jim bower [mailto:bower at uthscsa.edu] 
Sent: Tuesday, 29 July 2008 9:33 AM
To: r.gayler at gmail.com
Cc: comp-neuro at neuroinf.org
Subject: Re: [Comp-neuro] Review announcement

Ross

Great -  thanks for support from an unexpected source. ;-). 

While it is completely contrary to the intuition of most experimentalists
who are fond of their single neuron data, it is quite likely that individual
neurons in at least the mammalian brain don't matter. Although, as I have
mentioned before, one must be aware that neurobiologists pick stimuli to
maximize the presumption that they do. 

On the other hand, in the extreme one could find oneself arguing a la
Lashley that there is no structure in cortex and everything is completely
distributed. This is also clearly not true. I have suspected for some time
that the need for communication (as reflected very loosly in cortical
oscillations) is what detemines the size of the population of neurons across
which different types of computations are implemented but within which,
individuals actually don't matter much. These populations are probably
analogous to cortical areas. 

One other point however, is that abstract modelers tend to think of cortex
as one thing and so do many neurobiologists who like to think in terms of
things like the cortical microcircuit (or as in the case recently of the
blue brain project of a cortical column, which actually doesn't exist - but
anyway) in fact it is likely that cerebral cortex is a sequence of things.
Getting at what cortex does will require that we understand where this
sequence starts and ends. And, no time here and on my poor blackberry to
discuss this, but if you are trying to sort these things out in the context
of the visual system, as most are, you may be starting at the wrong end. It
seems to me likely that it is the olfactory system that invented basic
cortical circuitry. And there is a rather interesting story emerging there
about just how random cortical activity and connectivity is. 

Jim bower




More information about the Comp-neuro mailing list