Connectionists: The Atoms of Neural Computation: A reply to Terry Sejnowski

[Stephen Grossberg]

Dear Terry,

I personally don’t believe that a “debate between lumpers and splitters” is productive. Understanding the brain cannot be achieved in one grand step. It needs to be done by successive approximations, by unlumping coarser models into finer models. Such a method has been enormously successful throughout the history of theoretical physics.

In order to deeply understand how brains give rise to minds, I believe that we need to take seriously the basic fact that: Behavioral success drives brain evolution. This implies that, if we want to discover the evolutionary constraints that have shaped brain design, we need to theoretically represent the functional brain units that can compute indices of behavioral success. I have developed and practiced such a method, with many colleagues, over the past 50 years. It is summarized in http://www.cns.bu.edu/Profiles/Grossberg/GrossbergInterests.pdf.

That is how, after more than 20 years of developing non-laminar neural models of various behavioral functions, my colleagues and I were forced into finer laminar cortical models with a much expanded explanatory and predictive range. As one result, in Table 1 in Raizada and Grossberg (2003, Cerebral Cortex, 13, 100-113,  http://www.cns.bu.edu/Profiles/Grossberg/RaiGro03CerCor.pdf), the LAMINART model predicts distinct functional roles in  visual perceptual grouping for multiple identified cell types and connections throughout the layers of V1 and V2. Advanced experimental methods are needed to test these predictions, as well as to report unknown circuit properties, but I doubt that having such data without a functionally meaningful theory would make much sense.

As another example: the TELOS model (Brown, Bullock, and Grossberg (2004, Neural Networks, 17, 471-510, http://www.cns.bu.edu/Profiles/Grossberg/BroBulGro2003NN.pdf) learned five saccadic tasks that are reported in monkey experiments and then, using the learned model parameters, quantitatively simulated, and predicted the function of, the recorded dynamics of 17 identified cells types across multiple brain regions, as well as cell types that have not yet been reported. Again, more experiments are needed, but without a functionally meaningful theory, the data may well seem meaningless.

I could go on with many such examples, but I hope that my main point is clear: In order to develop this kind of evolving but principled theory, which embodies both "lumping" and "splitting", I believe that one needs to link brain to behavior at every step, and acknowledge that the theorist’s task is to provide functionally meaningful explanations and predictions for every component in a model, at each stage of the unlumping process, and at multiple levels of organization.

Best,

Steve

 
On Nov 3, 2014, at 5:13 PM, Terry Sejnowski wrote:

> The debate between lumpers and splitters on cortical areas will not be settled
> until we have the right tools to probe them anatomically and functionally.
> 
> We don't even know how many types of neurons there are in the cortex.
> Estimates range from 100 to 1000.
> 
> One of the goals of the BRAIN Initiative is to find out how many
> there are and how they vary between different parts of the cortex:
> 
> http://www.braininitiative.nih.gov/2025/index.htm
> 
> An important source of variability between neurons is differential patterns 
> of gene methylation, which is uniquely different in neurons compared 
> with other cell types in the body:
> 
> Lister, R. Mukamel, et al.  Global epigenomic reconfiguration 
> during mammalian brain development, Science, 341, 629, 2013
> 
> http://directorsblog.nih.gov/2013/08/27/charting-the-chemical-choreography-of-brain-development/#more-1983
> 
> http://papers.cnl.salk.edu/PDFs/Global%20epigenomic%20reconfiguration%20during%20mammalian%20brain%20development%202013-4331.pdf
> 
> We now have optical techniques to record from 1000 cortical neurons 
> simultaneously and that will increase by a factor of 100-1000x 
> over the next decade.
> 
> This will create a big data problem for neuroscience that readers of
> this list could help solve:
> 
> Sejnowski, T. J. Churchland, P.S. Movshon, J.A. 
> Putting big data to good use in neuroscience, 
> Nature Neuroscience, 17, 1440-1441, 2014
> 
> http://papers.cnl.salk.edu/PDFs/Putting%20big%20data%20to%20good%20use%20in%20neuroscience%202014-4397.pdf
> 
> Terry
> 
> -----

Stephen Grossberg
Wang Professor of Cognitive and Neural Systems
Professor of Mathematics, Psychology, and Biomedical Engineering
Director, Center for Adaptive Systems http://www.cns.bu.edu/about/cas.html
http://cns.bu.edu/~steve
steve at bu.edu




-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.srv.cs.cmu.edu/pipermail/connectionists/attachments/20141103/5be4dcd6/attachment.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: Reply to Gary Marcus via connectionists 11-3-14.pdf
Type: application/pdf
Size: 102628 bytes
Desc: not available
URL: <http://mailman.srv.cs.cmu.edu/pipermail/connectionists/attachments/20141103/5be4dcd6/attachment.pdf>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.srv.cs.cmu.edu/pipermail/connectionists/attachments/20141103/5be4dcd6/attachment-0001.html>