[Comp-neuro] Re: Attractors, variability and noise
james bower
bower at uthscsa.edu
Thu Aug 14 22:31:42 CEST 2008
The definition I proposed for reaslistic models previously was
agnostic with respect to the form of the models --yes we use HH
compartmental models at the single cell level, but we have used other
forms too - including even forms that are much more Bard like than
Bower like (c.f. Crook, S.M., Ermentrout, G.B., and Bower, J.M.
(1998) Spike frequency adaptation affects the synchronization
properties of networks of cortical oscillators. Neural Computation.
10:837-854.).
Realistic modeling is an attitude and an approach -- not a particular
form of model -- In some sense it is like pornography - you know it
when you see it. However, intentionally representing neurons in a
form that is convenient for the function you are imposing, is a dead
give away that one is involved in modeling as proof of concept, rather
than modeling as a tool to figure out what you don't know.
Ultimately, it is the brain itself that will dictate what tools are
appropriate, and the nature of the understanding that can be achieved.
In conversations with my friends who have stayed in "Physics", I have
often been asked why anyone would ever want to take on the study of
the brain, given the clear insuffiency of the tools available.
I sometimes answer that it is clear from the history of science that
new tools follow new problems. Often the response has been -- "sure
-- but why not study some intermediate system, like for example the
weather or ocean currents, or something that can only be done with
numerical simulations (rather than closed form solutions), but at the
same time can relay on physics we understand (thermodynamics,
turbulence, etc)." "Why try to make such a large step into the unknown"
Really, of course, the truth is that we are all interested in the
brain, and have been foolish enough to think that we can start to
study it now, absent the tools.
Given that, however, as I have said already in this discussion, one
has to be very careful about the extent to which available tools
dictate how you think about the system -- call this the "Tyranny of
the Tools".
I do think we can say something already about the kinds of tools we
need to continue to make progress - and I agree with several previous
comments, that those tools are all tied up with the technology of
numerical simulations, including, for example, mechanisms for crossing
scales, understanding complex parameter spaces, building the
simulations themselves, visualizing the results, comparing different
models and their performance, cross simulator operations,
parallelization, and importantly, using simulations themselves to
educate new generations of neuroscientists both in how to build
simulations and also about our current state of knowledge about
brains. This is why the Book of GENESIS was written in two parts, the
first part using GENESIS-based models to teach neuroscience and the
second part showing how those models can be changed to do further
research.
I would say, however, that one of the most important tool-based
innovations we need to move forward, is to replace our current methods
of publishing results with a form of publication appropriate for
neurobiology. It is no coincidence that Newton and his colleagues, at
the same time they were developing the theoretical basis for modern
physics, also invented the modern scientific journal. The problem is
that the form of that journal (a few pages, some figures, some
equations) is a completely inappropriate way to represent knowledge
about complex biological systems and realistic models in particular.
Neuro-DB is a step in the right direction, but only a baby step. With
Neuro-DB one can 'publish' the model on which a journal article is
published. However, what we should be doing is publishing the models
and building the text and figure descriptions of its behavior around
them. Why shouldn't figures for publication be generated from the
model itself? Why shouldn't textual descriptions be directly linked
to model components. And why shouldn't one be able to link to the
experimental datasets from which the model was tuned? Is there any
reason in principle, for example, that some of the peer review of such
publications couldn't be done automatically? What would that do to
the geo-political mess of peer review that we all now live with -
where some would say the tyranny of ideas has all but stifled
advancement in our science?
Once a model is submitted for publication, why can't we figure out how
to automatically test it for important properties like: 1) how
fragile are the principle results with respect to key parameters; 2)
how accurately does the model replicate the data? 3) does one set of
parameters REALLY produce all the results? One could even imagine
some formal (?Baysian?) basis for a comparison between data and
modeling results, or even determining whether a new model is
significantly better or even different from prior models. (cf. see
Baldi, P., Vanier, M.C., and Bower, J.M. (1998) On the use of
Bayesian methods for evaluating compartment neural models. J.
Computational Neurosci. 5: 285-314) Obviously, given the structure of
GENESIS (or Neuron), one could even, in principle, automatically track
the antecedents to the new model -- and start to apply a more rigorous
standard for individual contributions or the lineage of ideas.
I personally believe that this change in publication process will
eventually be critical for us to advance.
As some of you are aware, we have spent the last 5 years working on a
fundamentally new design for the GENESIS project -- we will shortly be
publishing several papers describing that design. We will also start
releasing beta versions for testing. Core objectives for the redesign
are: simulator interoperability, integrated multi-scale simulations,
and model-based publication. We are also currently planning a meeting
in San Antonio in March of next year, to work on practical aspects of
all of the above with other simulator and middle wear developers --
Stay tuned.
Jim
On Aug 13, 2008, at 6:39 PM, James A. Bednar wrote:
> | From: comp-neuro-bounces at neuroinf.org On Behalf Of jim bower
> | Sent: 01 August 2008 13:45
> | To: bard at math.pitt.edu
> | Subject: Re: [Comp-neuro] From Socrates to Ptolemy
> |
> | But, biology is different and the difference and the conflict is
> | perhaps best indicated in the difference between abstracted models
> | and "realistic" models.
> |
> | First, I would define realistic models not only as those that
> | include as much of the actual structure as possible, but also and
> | perhaps most importantly, models that are "idea" nuetral in their
> | construction.
>
> Hi Jim,
>
> Your written definition of a "realistic" model sounds reasonable to
> me, but in practice it seems like you only consider multicompartmental
> Hodgkin/Huxley models to be realistic (based on this discussion and
> your talks at previous GUM/WAM-BAMM meetings). Is that true?
>
> I agree that multicompartmental models are expressed at the right
> level for addressing questions about single neurons. However, I can't
> agree that such models are "idea neutral" or "not abstracted", because
> in practice they rely on an idea that the rest of the brain can
> reasonably be modeled extremely abstractly, e.g. as a set of spike
> trains following some simple distribution. Such an assumption only
> makes sense to me if the research questions are all about the behavior
> of that one neuron, rather than the coordinated behavior of the large
> populations of neurons that underlie functions like early vision in
> mammals (my own area of interest).
>
> For understanding visual processing, "including as much of the actual
> structure as possible" means including the hundreds of thousands or
> millions of neurons known to be involved. When it is not possible to
> build tractable multicompartmental models of that many neurons (due to
> the enormous numbers of free parameters and the computational
> complexity), there is no way to avoid relying on an idea or
> abstraction. One must choose either to simplify the neuronal context
> (by modeling only a few neurons at a very detailed level), or to
> simplify the models of each neuron. Someday perhaps we might not have
> to make such a choice, but certainly today and for the foreseeable
> future we do.
>
> Which option to choose, then, depends on the question being asked,
> which depends on the experimental techniques available and on the
> modelers' "ideas" and "abstractions" about what is important. If
> trying to understand and explain detailed measurements of single
> neurons, then sure, build multicompartmental models of single neurons,
> with an embarrassingly simplified representation of the rest of the
> brain (and of the environmental and behavioral context). But if
> trying to understand data from large neural populations, e.g. from
> 2-photon calcium imaging of patches of hundreds or thousands of
> visually driven neurons, or from optical imaging of even larger areas,
> then a "realistic" model should contain huge numbers of simpler models
> of the individual neurons, so that the large-scale behavior can be
> matched to the data available. Such forms of imaging data don't
> provide any meaningful constraints on the parameters of a
> multicompartmental model, but they do match very well to simplified
> single-compartment (point-neuron) models. Thus I would argue that
> very simple point-neuron models can be appropriate and "realistic", if
> they are embedded in a large network whose behavior is closely tied to
> imaging results at that level (as in my own current work).
>
> Of course, one could argue that large-scale imaging methods that only
> measure firing rates or average membrane potentials (and indirectly at
> that!) are simply irrelevant, precisely because they don't capture
> enough of the behavior of individual neurons to constrain a
> compartmental model. But that's an idea about what is important and
> what is ok to ignore, just like my own opinion that neural behavior
> only makes sense in the context of large populations.
>
> Jim
> _______________________________________________________________________________
> Dr. James A. Bednar
> Institute for Adaptive and Neural Computation
> University of Edinburgh
> 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 Neuroscience
Research Imaging Center
University of Texas Health Science Center -
- San Antonio
8403 Floyd Curl Drive
San Antonio Texas 78284-6240
Main Number: 210- 567-8100
Fax: 210 567-8152
Mobile: 210-382-0553
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://www.neuroinf.org/pipermail/comp-neuro/attachments/20080814/76288144/attachment-0001.html
More information about the Comp-neuro
mailing list