neural coding of behavior: evidence for precise timing of spikes ?

Mark Laubach laubach at biogfx.bgsm.wfu.edu
Thu Dec 7 12:51:27 EST 1995


jan at neuroinformatik.ruhr-uni-bochum.de (Jan Vorbrueggen) wrote:

>In article <49tt44$otg at eis.wfunet.wfu.edu> laubach at biogfx.bgsm.wfu.edu (Mark
>Laubach) writes:

>   My own search of the literature indicates that there is no evidence for
>   precise timing of spikes as a neural code in awake, behaving subjects.

>Besides Abeles, whom you mention below, there is also recent work from
>Singer's lab in Frankfurt using behaving animals.

Could you please comment on this work?  The presentation I saw at this
year's SFN meeting was about local field potentials, not single-unit
activity.

>   Rather, those who have searched for such coding have instead found that
>   local changes in firing rate, not the precise temporal pattern of spikes,
>   may serve as a code for environmental stimuli, movements, task
>   contingiencies, etc. (e.g., Richmond's work on visual cortex).

>I always attribute such results to overtraining: if your animal has spent half
>a year to a year on learning a very specialised, narrowly defined task, it
>will develop a specialised (group of) cell(s) subserving that task, and the
>flexible mechanism for temporal binding is no longer required. See von der
>Malsburg's recent article in Current Opinion in Neurobiology.

Actually, the animals (rats) used in my task are trained in 2-6 weeks
to perform a simple reaction-time task (i.e., sustain a nose-poke for
500-2000 msec until tone).  Overtraining is only for one week.  

I agree with your point for animals that are trained for months.  For
such studies, one has to ask if the behavior studied is really the
behavior of interest, as the task becomes "habitual" for the animal.

Nevertheless, the task must be "narrowly defined" if one is to make
anything of the neural activity and for more complex tasks one has to
overtrain.

With regards to the ever sexy term of "binding", we have found in fact
evidence for "temporal binding", i.e., bilateral synchronization of
spike activity in the striatum at particular points in time in
relation to performing the task (see SFN abstract for 94: SF Sawyer).
However, this synchrony is not on the order of several milliseconds
but instead on the order of 10s of milliseconds.  Also, you have to
use an appropriate method to see these correlations, i.e., JPSTH.
(Paper should be together by this Spring.)

In any case, the question is "How does a correlation in activity
between cells or an observed repeating pattern of spikes represent a
neural code?"  This seems to require a _statistical_ procedure to
prove that the pattern of activity is _predictive_ of the animal's
behavior or level of performance.  Otherwise, the data are simply
descriptive.  With regard to a correlation, I do not see how one can
use a measure based on an overall relationship to get at
trial-by-trial variation that is the basis for prediction.

>   I know that some (e.g., Abeles) have reported that precise spike patterns
>   across small ensembles of neurons can occur in behaving subjects, but have
>   these patterns been shown to "be good for anything" with regard to the
>   subject's performance of the task?

>Firstly - what more can you, realistically, want from an experiment using whole
>brains? Secondly - these temporal relationships are highly reproducible to the
>millisecond level (while some dispute that nerve cells are even able to do
>this, e.g. Newsome et al. ...); Occam's razor tells you that evolution didn't
>implement this by accident.

By "good for anything" I meant  to ask:  "Do the patterns occur in a
reproducible fashion that would allow one to predict what happened in
the task from trial to trial?"  What I want is the ability to
recognize trials when the animal's performance varied using only spike
acitivty of single cells and of neural ensembles.  In fact, I can do
this!  (and so can others: see Deadwyler and Hampson's review in last
week's Science)  The goal here is to understand how neural ensembles
process information and to build a basis of knowledge (i.e., identify
features in patterned neural activity) for constructing devices that
could "read the brain" (i.e., neural protheses).  

Moreover, exact temporal patterns of spikes from a single cell or
within an ensemble, my data from cortex and striatum, are NOT
reproducible from trial to trial.  Such patterns may occur more than
is expected by chance.  The question is "Can the patterns be used to
infer something about the behavior of the animal?"  You can not answer
this question in a dish or with electrical stimulation!

I challenge those who claim to have found such "meaningful patterns"
to examine the statistics of these patterns as predictors for
behavior.  I also ask them to use _appropriate methods_ for this: for
example, nonlinear signal processing (e.g., wavelets) and
nonparametric classification (e.g., CART or backprop).  (A paper on
these methods should be together by the end of January.)

Finally, I agree with the conclusions of Newsome et al.  My own
studies of pattern recognition using neuron-like  Poisson processes
shows that such processes are NOT entirely reliable transmitters of
information, if information means firing a pattern of spikes that
repeats from observation to observation.  In fact, such processes
_fail_ to fire on some occasions, even when the local rate is greater
than the ever sexy 40 Hz!


Mark Laubach




More information about the Neur-sci mailing list