Connectionists: "Abstract" vs "Biologically realistic" modelling

Tue Jan 28 23:26:57 EST 2014

Jim, 
With all the talk of physics, let's not forget Feynman's eloquence in explaining an unsolved and old problem in physics:  turbulence.  Where the understanding of the parts does not lead up to much of an understanding of the whole.  

"How vivid is the claret, pressing its existence into the consciousness that watches it!  If our small minds, for some convenience, divide this glass of wine, this universe, into parts--physics, biology, geology, astronomy, psychology, and so on--remember that nature does not know it!  So let us put it all back together, not forgetting ultimately what it is for.  Let it give us one more final pleasure:  drink it and forget it all!"

								-Richard Feynman, from Six Easy Pieces

On Jan 28, 2014, at 10:07 PM, james bower wrote:

> 
> On Jan 28, 2014, at 7:02 PM, James A. Bednar <jbednar at inf.ed.ac.uk> wrote:
> 
>> |  Date: 2014-01-25 12:09 AM
>> |  From: james bower <bower at uthscsa.edu>
>> | 
>> |  Having spent almost 30 years building realistic models of its cells
>> |  and networks (and also doing experiments, as described in the
>> |  article I linked to) we have made some small progress - but only by
>> |  avoiding abstractions and paying attention to the details.
>> 
>> Jim,
>> 
>> I think it's disingenuous to claim that *any* model can avoid
>> abstractions,
> 
> First, you might want to read the paper I linked to earlier, so that you understand what I am saying.
> 
> here it is again:  https://www.dropbox.com/s/r046g03w8ev5kkm/272602_1_En_5%20copy.pdf
> 
> With respect to that article, here is a list of properties of Purkinje cells that if you don’t have them, means you aren’t modeling a Purkinje cell, based on 40 years of modeling and experimental studies not only including our own:
> 
> - no dendrite
> - a dendrite when present that simply does a voltage sum
> - or in other words a dendrite with no active conductive properties
> - a soma that has a simple fixed defined threshold to fire
> - or in other words, a dendrite with no active conductive properties
> - a dendrite that isn’t morphologically based on an actual reconstructed Purkinje cell dendrite.
> - a soma that doesn’t generation action potentials independently on its own
> - a dendrites whose principe effect is to stop the soma from firing
> 
> As far as its response in the context of a realistic network model:
> 
> - a dendrite that only receives excitatory input, without inhibitory input
> - input from parallel fibers directly driving somatic spiking
> 
> 
> These are equivalent, or should be, to making a model of the interaction of subatomic particles and ignoring conservation laws.
> 
> You couldn’t get away with it in physics - happens all the time in modeling Purkinje cells and the cerebellum (even last week).
> 
> Neuroscience should be accumulating these biological equivalents to the laws of physics - but it isn’t.  (would even be useful to the Neural Network, connectionists, machine learning community who mostly have a pre hodgkin huxley view of neurons).
> 
> But there is another and more definition, in my view, of ‘realistic’ models, which again has to do with process.  
> 
> The vast majority of neurobiological models are designed to demonstrate that an idea somebody had about how the system worked is plausible.  No problem with that, what so ever, in engineering (its what engineers do) - but a significant problem in ‘reverse engineering’ the brain.  
> 
> Realistic models, in my nomenclature (and I was one of the first people to use the term actually) aren’t defined explicitly by how adorned they are or not with biological stuff.  They are models whose construction and parameter tuning is primarily and fundamentally aimed at replicating basic biological data.  Not, synthesized biological data (ie. ocular dominance columns or orientation selectivity), but basic recorded responses that don’t have known functional implications.  Better yet, biological data recorded under completely artificial circumstances and conditions which never-the-less reveal complex behavior that isn’t understood.  In the case of realistic single cell models, that data is often from voltage clamp experiments.  Clamping the voltage of a neuron at a fixed level by injecting current is a highly artificial thing to do - yet, in many cells and in particular the Purkinje cell, it reveals a complex pattern of activity reflecting in a complex way the biophysical structure of that neuron.  Once model parameters have been tuned to replicate voltage clamp data - then, one freezes these parameters and applies synaptic input to see if one can replicate the basic response properties of the cell (e.g. its variable rate of action potential generation).  In our experience at that point we have always started to find all sorts of things that you didn’t know where there.  In the case of the PUrkinje cell, for example, we found out that it didn’t matter where on its huge dendrite you applied a synaptic input, that input had the same effect on the soma.  If you want to know why that is interesting (and what happened next) take the time to read the paper.
> 
> So, the point is this - again, the real litmus test for realistic modeling, is whether the model was tuned and designed to produce a particular functional result - and then adorned with biology, of if replicating the basic biology independent of function was the first step and remains the reference step for the modeling work.
> 
> Again to return to Newton - he apparently built a ‘realistic’ (by my definition) model of the moon orbiting the earth.  he applied mathematical analysis to figure out the size of the force holding it in its orbit.  He then realized that that force appeared to be the inverse square of the distance.  Actually, the force he calculated the first time wasn’t - it was less - and therefore, Newton at age 19, apparently thought that there was some other force (Keplers vortex force), in the mix as well.  It wasn’t until many years later, after turning most of his attention to alchemy, when informed that another scientist was about to report the inverse square relationship that he became interested again - that interest turned into the work that ended up with his treatise on mechanics.
> 
> So, the bottom line - if you want to understand the nervous system and you have an idea about how the cerebellum is involved in learning - and you build a model that implements your scheme - then, you are engaged in a Ptolemaic effort, not a realistic one.  If most of the “predictions’ of your model are actually “postdictions’ of well known phenomena that you are using to convince people your functional idea is right, then again, you are in the domain of Ptolemy.  If you are building a model to solve the traveling salesman problem, or to perform better voice recognition on a chip - good for you - that’s how engineering works.  No problem with that at all - and actually, if you picked something up from a neuroscience lecture that gave you a new idea about how to make your neural networks chip - absolutely no problem what so ever with that either - as we all know, biology has served as an important source of creativity in engineering historically.  However, if you want to claim that your model also reveals something important about how brains work, then the model must either be ‘realistic’ first, or be able to link to such a model.
> 
> It goes without saying that these types of realistic models can be built at many levels, as long as the model has biological components.  (you won’t convince me with mean field theories of cerebral cortex).  It also goes without saying, of course, that we don’t have the technology or the knowledge for that matter to build one model of everything - although personally, I believe eventually we will have to, and reflecting that view, Version 3.0 of GENESIS was specifically built to link broadly across many different levels of scale.
> 
> The critical question therefore, is whether the model is built in such a way that the biology can tell you something you didn’t know before you started (just like the earth moon model told Newton) - or, is the biology just dressing up something you already believed to be true and just wanted to convince the rest of us.  Building the model out of realistic components, and then testing it on theory- neutral biological data, is more likely to lead to the former.  At least it has over and over again for us.
> 
> 
> Jim
> 
> 
> 
>> and in particular that your type of "realistic"
>> multicompartmental single-cell and network modelling could ever do so.
>> 
>> *Real* morphologically complex cells are embedded in complex networks,
>> which are embedded in complex organisms, which are embedded in complex
>> environments, which are embedded in complex ecosystems.  Evolution
>> acts on the net result of *all* of this, indirectly via a process of
>> development.  Certain species thrive in certain ecosystems if their
>> proteins, cells, networks, nervous systems, bodies, and communities
>> allow them to function in that environment well enough to reproduce.
>> The details of *all* of these things matter.
>> 
>> Are all of these details represented realistically in your models?
>> No, and they shouldn't be -- you pose questions that can be addressed
>> by the things you do include, abstract away the rest, and all is well
>> and good.  But other different yet no less realistic models are built
>> to address different questions, paying attention to different sets of
>> details (such as large-scale development and plasticity, for my own
>> models), and again abstract away the rest.
>> 
>> I am happy to join with you to decry truly unrealistic models, which
>> would be those that respect none of the details at any level.  Down
>> with unrealistic models!  But there is no meaningful sense in which
>> any model can be claimed to avoid abstraction, and no level that
>> exclusively owns biological realism.
>> 
>> Jim Bednar
>> 
>> ________________________________________________
>> 
>> Dr. James A. Bednar
>> Director, Doctoral Training Centre in 
>> Neuroinformatics and Computational Neuroscience
>> University of Edinburgh School of Informatics
>> 10 Crichton Street, Edinburgh, EH8 9AB  UK
>> http://anc.ed.ac.uk/dtc
>> http://homepages.inf.ed.ac.uk/jbednar
>> ________________________________________________
>> 
>> -- 
>> The University of Edinburgh is a charitable body, registered in
>> Scotland, with registration number SC005336.
>> 
> 
>  
> 
>  
> 
> Dr. James M. Bower Ph.D.
> 
> Professor of Computational Neurobiology
> 
> Barshop Institute for Longevity and Aging Studies.
> 
> 15355 Lambda Drive
> 
> University of Texas Health Science Center 
> 
> San Antonio, Texas  78245
> 
>  
> Phone:  210 382 0553
> 
> Email: bower at uthscsa.edu
> 
> Web: http://www.bower-lab.org
> 
> twitter: superid101
> 
> linkedin: Jim Bower
> 
>  
> CONFIDENTIAL NOTICE:
> 
> The contents of this email and any attachments to it may be privileged or contain privileged and confidential information. This information is only for the viewing or use of the intended recipient. If you have received this e-mail in error or are not the intended recipient, you are hereby notified that any disclosure, copying, distribution or use of, or the taking of any action in reliance upon, any of the information contained in this e-mail, or
> 
> any of the attachments to this e-mail, is strictly prohibited and that this e-mail and all of the attachments to this e-mail, if any, must be
> 
> immediately returned to the sender or destroyed and, in either case, this e-mail and all attachments to this e-mail must be immediately deleted from your computer without making any copies hereof and any and all hard copies made must be destroyed. If you have received this e-mail in error, please notify the sender by e-mail immediately.
> 
>  
> 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.srv.cs.cmu.edu/pipermail/connectionists/attachments/20140128/f70fa5da/attachment.html>