[Comp-neuro] Re: Attractors, variability and noise
Antonio Carlos Roque da Silva Filho
antonior at ffclrp.usp.br
Fri Aug 15 20:59:47 CEST 2008
Jim and others,
Regarding analogies, some of you may find interesting my own analogy between
brain models and artistic styles.
Artists (painters, sculptors, writers, photographers, flm directors, etc)
have their own way to "model" human nature (a consequence of the human
brain) and they have been doing this for much longer than computational
neuroscientists, so we may learn something from them. They also debate
strongly on the appropriate level of approaching nature, and their works are
usually classified as impressionistic, expressionistic or realistic.
I suggest that brain models also can be classified into this tripartite
Impressionistic neural models would be those which use reduced neuron
models, i.e. based on a small number of parameters. Such models convey the
"impression" of reality. Some examples (with varying impressionistic
degrees) are models that use single neuron models like Hindmarsh-Rose's,
FitzHugh-Nagumo's, Izhikevich's, integrate-and-fire, cellular automata,
McCulloch-Pitts, etc. These models attempt at an understanding (in a
dynamical systems sense) of the system under study. So far, they have been
very successful in providing such a kind of understanding for single neuron
spiking behavior but their usefulness for large-scale brain models (other
than speeding up simulations) has yet to be shown.
Expressionistic neural models would be those that "express" their authors'
views on the brain. They start from a number of a priori ideas about the
brain and, ideally (but not always), should be constructed so that these
ideas are explicitily present in the models so that they can be tested
experimentally. The most famous example of such models is Freud's
psychoanalysis (btw, his first attempt was to construct a biology-based
model but due to lack of data at
his time he ended up turning to the other end of the spectrum) but one also
may include in this class models based on hypotheses like Hebb's cell
assemblies, synfire chains, Freeman's K sets, chaotic itinerancy, etc. The
main characteristic of these models (for the good and the bad as has been
extensively discussed in this list) is that they propose interpretations of
the brain functions.
Realistic neural models would be those that attempt to reproduce structural
and physiological properties of cells and circuits in an accurate way.
Examples are HH-type models, compartmental models, etc. In principle, they
should be idea-free, as Jim pointed out, so that they place value on the
available experimental data. To put it another way, they are constructed and
can be used to test how good is one's choice of data to reproduce a given
behavior. The appeal of this kind of modeling is that, given the high degree
of brain complexity which exceeds our capacity of understanding it, they may
be the only option ahead. But this is not to be considered a drawback. The
demand for realism in art led to a lot of useful knowledge like, e.g.,
better anatomical descriptions, the laws of perspective, lighting techniques
and representation of movement (think of modern animation films). In film,
to cite another art analogy, several directors agree that the ideal to be
achieved is pure realism, like in a documentary: it would be shot without
any a priori explanatory attempt but, after being seen, would have explained
PS. Jim, thank you for publicizing LASCON. I wish I had the means to bring
some of the guys who have been discussing here to have an open discussion on
these issues with our students in 2010.
Departamento de Fisica e Matematica
FFCLRP, Universidade de Sao Paulo
14040-901 Ribeirao Preto-SP
Brazil - Brasil
Tels: +55 16 3602-3768 (office);
+55 16 3602-3859 (lab)
FAX: +55 16 3602-4887
E-mails: antonior at neuron.ffclrp.usp.br
antonior at ffclrp.usp.br
----- Original Message -----
From: "james bower" <bower at uthscsa.edu>
To: "Brad Wyble" <bwyble at gmail.com>
Cc: "CompNeuro List" <comp-neuro at neuroinf.org>
Sent: Wednesday, August 13, 2008 4:23 PM
Subject: Re: [Comp-neuro] Re: Attractors, variability and noise
>I would rather personally gain insights from including a known feature of
>the brain in a model than randomly misplacing parenthesis - :-)
> perhaps a new application of genetic algorithms here - with respect to
> source code for abstract models. :-)
> However, the general point is absolutely taken -- without something
> concrete and mathematical, you don't know what you know or what you don't
> know. Further, unless you share your model with others (and I don't mean
> through paper publication), they don't know what you know, they know, or
> collectively you don't know either.
> Another problem with abstract models -- most of which are simple enough
> that you can write your own code. Systems like Bard's XPP are absolutely
> essential to have some form of intercommunication -- and BTW, what about
> misplaced parenthesis that go unrecognized?
> Often when I talk to biologists about the need for modeling, they tell me
> that they don't yet know enough to build a model - truth is, you don't
> know how little you know until you start to build one (I may have already
> said that).
> For sure I have said before (several times in several different ways)
> that a model should NEVER be principally designed to prove to people you
> are right (smart, sophisticated, or etc). Unfortunately, many are.
> On Aug 13, 2008, at 1:46 PM, Brad Wyble wrote:
>> At the risk of missing my flight I can't resist continuing this debate.
>> As I understand the other end of the spectrum, we construct
>> increasingly realistic models and end up with a simulated brain without
>> a real understanding of how it works, which makes no sense to me.
>> Understanding is what we're after, and that understanding can only
>> reside in the brains of the population of scientists, not in their
>> I suspect that I have created a straw man here, but I'm curious to what
>> extent I've abused your position.
>> Haven't abused at all -- with one big exception -- realistic models are
>> more likely to tell you how things work, than are models in which 'how
>> things work' is assumed. In our experience, realistic modeling has
>> consistently and steadfastly told us things that we didn't know before -
>> problem is, those things fly in the face of many of the current
>> 'theories" operating in the parts of the brain I study, making the
>> publication of papers, getting grants, etc, much more difficult.
>> BTW, almost every time, the models have also made it clear that I was
>> wrong in how I was thinking about the system:
>> I agree wholeheartedly but abstract models are just as capable of
>> telling us new things. To cite a specific example of my own, my current
>> modelling effort (which explains a quite high-level phenomenon of visual
>> attention) features a recurrent excitation between targets and
>> attention that I initially implemented by misplacing a parenthesis in an
>> equation. I realized quite quickly that this circuit worked better
>> than the one I had intended to create, and is just as plausible, if not
>> more so, than what I had in mind.
>> So a major contribution of models is to allow us to explore the behavior
>> of systems more complicated than we can reason about in our heads. And
>> it turns out that human reason hits its limit quite quickly; even a
>> model with a handful of abstract, rate-coded neurons is informative in
>> this respect.
>> As for the holy grail of a realistic model of the entire brain, is there
>> such a thing as enough detail?
>> I think that if a time traveller from the future dropped a simulation of
>> the brain, realistic down to the level of RNA synthesis, in our laps,
>> many of the realists would want to continue drilling down to the sub
>> molecular level and we'd be having the same debate all over again.
>> The rest of us would start trying to build abstract theories on top of
>> this simulation, so I think we might as well get started with what we
>> already know.
>> There is hope
>> Yes, however I think you have succesfully highlighted some glaring
>> difficulties with the way our discipline is currently running. I
>> think the way out is not to focus on a particular end of the realism/
>> abstract spectrum, but to do a better job of avoiding the tyrannical
>> ideas by focussing on data-driven theory.
>> Hrmph, I think I have ended my short contribution to this debate back
>> where we started from.
> 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
> e-mail in error or are not the intended recipient, you are hereby
> that any disclosure, copying, distribution or use of, or the taking of
> action in reliance upon, any of the information contained in this e- mail,
> any of the attachments to this e-mail, is strictly prohibited and that
> 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
> your computer without making any copies hereof and any and all hard
> made must be destroyed. If you have received this e-mail in error, please
> notify the sender by e-mail immediately.
> Comp-neuro mailing list
> Comp-neuro at neuroinf.org
More information about the Comp-neuro