Connectionists: Real or virtual? Grandmother cells or "virtual" grandmother patterns?

[Tim Shea]

Hi Richard,

I'm sorry to say I haven't ready your linked paper, though I will soon.
Jeff Rodny, Chris Kello, and I have a paper in a special issue of *Language,
Cognition, and Neuroscience* on the topic of evidence for and against
grandmother cells. The issue was organized by Jeff Bowers, and he wrote a
very helpful summary of the contributions here:
http://www.tandfonline.com/doi/full/10.1080/23273798.2016.1267782.

In our paper, we discuss evidence and provide a computational model as a
proof of concept to argue that single unit responses (whether localist or
distributed) may be transient across various timescales without affecting
"behavior". I wonder how well this maps onto your concept of a virtual
pattern? The basic insight we offer is that, to the extent neurons and
neural systems are plastic, relationships between external features and
single unit activity may be metastable. Brief snapshots of these
relationships can exhibit a high measured stability, particularly for a
subset of neurons selected *because *of their stable responses during the
recording window. Thus, 30% of neurons may be highly selective for 30
minutes, 10% may be highly selective for 2 hours, etc.

That being said, we present this side of the argument as a counterpoint to
the ongoing debate over distributed and localist representations, in which
both sides explain only the static properties of the representations.
Neither my coauthors nor I feel that all representations in the brain are
likely to be so transient, and even if this were true it would present a
serious challenge to what is even meant by representation (not a challenge
that we are eager to tackle).

Your thoughts and feedback are welcome on the paper:
http://www.tandfonline.com/doi/abs/10.1080/23273798.2016.1242760
PDF link <http://cogmech.ucmerced.edu/pubs/RodnyETAL2016-LCN.pdf>

Best,
Tim Shea
University of California, Merced

On Sat, Aug 26, 2017 at 8:54 PM, Richard Loosemore <rloosemore at susaro.com>
wrote:

>
> Back in 2010 I wrote a paper with T.A. Harley in which we argued that
> "concepts" (like [grandmother]) are likely to be *virtual* patterns of
> activity on neural circuits and NOT hard coded into the circuits themselves.
>
> So, for example, a virtual pattern of activity might come and go between
> active and dormant states, or it could move around the brain (perhaps from
> one column to another).
>
> In support of this conclusion, we pointed out that  common interpretations
> of brain imaging data were simply not consistent with the usual assumption,
> which is that neurons directly represented concepts.  We pointed, in
> particular, to the infamous Jennifer Aniston Cell paper, where the virtual
> concept hypothesis was the only viable one.
>
> And yet, the "virtual" idea is almost completely absent from the
> literature.  Why?  If concepts are virtual, this would make a nonsense of
> many interpretations of neuroscience results, because firing patterns would
> only have a weak relationship to meaningful entities like concepts.  (Think
> about it:  if concepts can wander around the cortex, what is the point in
> saying that a particular place in cortex corresponds to a semantically
> tangible thing?).
>
> Anyway, I note that a recent paper from Laura N. Driscoll, Noah L. Pettit,
> Matthias Minderer, Selmaan N. Chettih, and Christopher D. Harvey (Dynamic
> Reorganization of Neuronal Activity Patterns in Parietal Cortex):
>
> http://www.cell.com/cell/pdf/S0092-8674(17)30828-0.pdf
>
> (Overview here: http://news.harvard.edu/gazette/story/2017/08/brain-
> flexibility-changes-the-way-we-remember-and-learn/)
>
> ... brings yet another confirmation of the "virtual" concept hypothesis.
> The most parsimonious interpretation of their results is that the activity
> patterns are changing precisely because the "concepts" (when active) are
> not identifable with fixed hardware, but are actually virtual.
>
> It seems to me this is one of the most important issues in all of
> neuroscience, since it changes the flavour of every result out there.
>
> What do you think?
>
> ---
>
> Richard Loosemore
>
>
>
> Reference
>
> Loosemore, R.P.W. & Harley, T.A. (2010). Brains and Minds: On the
> Usefulness of Localization Data to Cognitive Psychology. In M. Bunzl & S.J.
> Hanson (Eds.), Foundational Issues in Human Brain Mapping.Cambridge, MA:
> MIT Press
>
> https://www.academia.edu/563588/Brains_and_Minds_On_the_Usefulness_of_
> Localization_Data_to_Cognitive_Psychology
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.srv.cs.cmu.edu/pipermail/connectionists/attachments/20170828/889c6de0/attachment.html>