[Comp-neuro] Re: (1) Maybe channel biophysics is not the "real"
level for analysis of brain function.
Harry Erwin
harry.erwin at sunderland.ac.uk
Fri Aug 22 10:06:04 CEST 2008
On 21 Aug 2008, at 23:30, james bower wrote:
> So, how appropriate really is this statistics to a device whose
> dynamical behavior is "held in the biochemical states of its
> neurons, developed through environmental experience, as they behave
> in complex, dynamically shifting networks." And how much is this a
> leap of faith?
>
> Which leads me to the following bit of history -- as a postdoc, I
> developed a technique (in Llinas' lab) to record from multiple (32)
> cerebellar climbing fibers at once -- I took that data to Wisconsin
> and to a wonderful mathematician named Josh Chover, and asked him
> how the heck to analyze this data. He and I ended up teaching a
> course (in 1982) on multi-neuron data analysis (A guy named Matt
> Wilson was the TA). During the course, i realized that not only
> PSTHs but also correlation analysis (still the mainstay of this type
> of data presentation), was completely inappropriate - Despite taking
> multiple statistical courses as an undergraduate, I also finally
> realized the critical connection between the assumptions of the
> model from which any particular statistics is constructed (point
> process statistics) , and the organization of the system to which
> they were applied (neuronal firing patterns). Faced with the
> realization that point process statistics was completely
> inappropriate to analyze neuronal firing patterns - I asked myself,
> "what is the appropriate model from which to extract statistics to
> study the brain?" -- I decided that, in the limit, that model was
> the brain itself -- which is the original origin of my interest in
> building realistic models. I convinced Matt Wilson to build a model
> of the olfactory cortex (over my supervisor Lew Haberly's
> objections), Matt came to Caltech where I bullied him into building
> the framework for GENESIS (to which he has never returned) -- and
> the rest, as they say is (if we are lucky) history.
>
> So, I actually believe that the process of building analytical
> infrastructure around realistic models, is really the process of
> building statistical devices to keep shaping the models based on
> biological data. And, I am suspicious of any level of modeling or
> analysis that relies on more generalized statistical models (like
> point process statistics).
>
> Jim
I had a similar, relevant experience last year. The following is from
a draft of a paper I was working on at the time:
"One goal of the MiCRAM programme is to develop a biologically
plausible model of auditory processing in the frequency lamina of the
inferior colliculus (IC) to clarify the roles of the multiple spectral
and temporal representations present at that level, The first step
towards that goal is the development of biologically realistic GENESIS
(M. Wilson et al., 1991) models of the individual disc cells of the
IC. That step involves exploring the development of a statistically
valid model of the firing patterns of these cells.
"A number of research groups are doing research into quantitative
models of neural spike trains (E. N. Brown et al., 2001; R. E. Kass
and V. Ventura, 2001; M. C. Wiener and B. J. Richmond, 2002; E. N.
Brown, 2003; M. C. Wiener and B. J. Richmond, 2003; U. T. Eden et al.,
2004). Since spike trains are essentially binary in form and vary
depending on the stimulus, evaluating goodness of fit has been found
to be more challenging than for continuous processes, and this is
particularly problematic for histogram-based models and rate measures
based on smoothed spike trains (E. N. Brown et al., 2001). To be able
to use standard statistical approaches to the problem, the data must
be transformed into a form that allows the use of distance measures
(E. N. Brown et al., 2001). Although modelling a neuronal spike train
as a Poisson process is simple and attractive, it is known to be
unrealistic (E. N. Brown, 2003), and more complex models are required.
"Kass and Ventura (R. E. Kass and V. Ventura, 2001) approached this
problem by modelling spiking probability as a function of the time
since stimulus onset and the time since the last spike. This results
in a class of inhomogeneous Markov interval (IMI) models. In a study
of neural spike data from the supplementary motor field of monkey,
they showed an IMI model outperformed an inhomogeneous Poisson (IP)
model (where both models were fitted using maximum likelihood) in
modelling the response of the neuron. Brown, et al. (E. N. Brown et
al., 2001) compared the fit of IMI, IP, and PSTH (peristimulus time
histogram, an inhomogeneous Poisson model fit to the instantaneous
rate function) models to the monkey data and showed Kass and Ventura’s
IMI model performed best and the IP model was better than the PSTH
model. The better performance of the IMI model reflected its explicit
treatment of temporal dependence in the spike train.
"Unfortunately it has been discovered that although IC disk cells fire
in a reliable way to repeated presentations of a constant intensity
stimulus, the same cell will respond quite differently if the
intensity of the stimulus is changed, which makes it difficult to use
quantitative models to classify these cells robustly (A. Rees et al.,
1997)."
Basically, I came to similar conclusions as JB. What was a bit amusing
was that other workers seemed more interested in following up my
citations than paying attention to my (negative) conclusions about the
usefulness of the approach.
--
"All ... pleas of convenience, even if their factual base is sound,
are inadmissible in principle." (Russell 1993)
Harry Erwin
More information about the Comp-neuro
mailing list