[Comp-neuro] re: realistic neuron models, challenges, Turing test and complexity

Wulfram Gerstner wulfram.gerstner at epfl.ch
Mon Aug 25 10:06:19 CEST 2008

Dear Bard, Carson, and Jim

you have all three  been talking about posing
challenges (Turing Test) for judging the quality of neuron models. The 
big question
is then: is a detailed biophysical model or a simplified model
going to win?

We actually have posed two challenges along these lines
in 2007 and 2008 and a new challenge will follow in 2009.
The task is to predict the behavior of a real neuron during new
experimental trials given the data collected with the SAME neuron
in previous experiments.  Some of the experimental data is derived from 
single electrode
conductance injection, other data is derived from  injection with two or 
electrodes simultaneously.

In all tasks, the aim was to quantitatively predict the behavior of the 
very same
neuron for which the earlier data used for parameter fitting was collected,
as opposed to predicting the qualitative behavior of a class of neuron
So, who wins?

For the experimental data we have and amongst all the submissions we 
the simple models win.  In principle I would expect that  eventually the 
biophysical models overtake, at least for the
three-electrode recordings,
but it seems that  the groups working with detailed models are very 
hesitant to submit:
Is this so because they are afraid to loose, or because they cannot
handle the large number of free parameters they have in their models?
I don't know.

Note that, in contrast to the discussion on this mailing list  the 
appropriate Turing test
is NOT to take the detailed model as a gold standard to which simple 
neuron models should compare.
Rather the gold standard is the EXPERIMENTAL neuron.
A model cannot be better than the neuron itself - and the neuron itself
shows fluctuations during the course of an experiment of, say, half an 
This limits the precision a model needs to have.

Of course, a simplified neuron model cannot account for ALL possible 
manipulations, for example pharmacological blocking of specific channels 
is not possible
in the model. However, one should require  that a simple neuron model  
for the type of stimulations a real neuron would receive in vivo - and 
this is, as you know, an
interesting subject in itself.

here the link:

best regards

Wulfram Gerstner
Professor for Computer Science and Life Sciences
Laboratory of Computational Neuroscience
Brain Mind Institute
Lausanne, Switzerland

More information about the Comp-neuro mailing list