Codes and time constants

[eplunix!peter@eddie.mit.edu]

Regarding mean rate vs. temporal codes, Bill Skaggs (4/4/95) commented:

> Instead of asking whether we are looking at a mean rate code or a
> temporal code (which is a meaningless question), we should ask what the
> shape of the time shift-vs-effect curve is, and in particular, what the
> largest and smallest time constants are.  Note that, although the shape
> of the curve may change if the effect is quantified in a different way,
> the time constants are likely to remain similar.

I know this is the way that many people think of the problem of 
rate vs. temporal codes, but it can lead to the conflation of codes and 
concepts which, in my opinion, really are different and should be kept distinct.

The issue of time constants is related to the temporal precision needed to
convey information via some code (what distortions in spike time pattern are
sufficient to change one message into another?). The issue of what spike
train patterns convey the message is complementary to the issue of precision.
(e.g. the average power of a signal is a different property than its 
Fourier spectrum, regardless of what sampling rate is used to specify the 
signal.)

An average rate code means that the average number of spikes within a given
temporal integration window is all that counts in determining the message,
i.e. rearranging spike time patterns without changing the number of spikes
within a window should have no effect (otherwise we would have something
more elaborate than a pure mean rate code). 

A temporal code is one in which the message sent is determined by:
1) the time patterns of spikes (e.g. the complex Fourier spectrum 
of the spike pattern) or 2) the particular spike arrival times relative 
to some reference event (e.g. the return time after an echolocation call, 
or absolute time-of-arrival relative to that of other spikes in 
spike trains produced by other neurons -- interneural synchrony).

For a temporal pattern code, if the time patterns of spike arrivals 
are scrambled without changing the mean rate, then the message is altered.
In the Covey/DiLorenzo electrical stimulation experiments in the 
gustatory system that I cited in the previous message, a particular time 
pattern of electrical pulses evokes behavioral signs in a rat that
it tastes a sweet substance, whereas scrambling the patterns while
maintaining the average number of pulses evokes no such signs. The gustatory
system is slow, so the temporal precision of the code is probably in the
tens of milliseconds, but nevertheless, the time pattern does appear to
be the coding vehicle, since its disruption evidently has perceptual 
consequences. 

The differences between rate-based and timing-based codes can also be seen
from the perspective of the decoding operations required of each. 
The neural operations needed to interpret rate codes are rate-integration
processes (all other things being equal, the longer the window 
the higher the precision), whereas those needed to interpret 
temporal codes are coincidence and delay processes
(the shorter the coincidence windows and the more reliable the delay
mechanisms, the higher the precision). In my opinion, this is why the
discussion of whether most cortical pyramidal neurons are performing 
rate-integrations vs. coincidence detections (and yes, on what time scales
they might be doing these things) is so crucial. It might even be possible
for a given neuron to be doing both, albeit on different time scales,
since particular time patterns of coincidence can be embedded in spike trains
also driven by Poisson-like inputs (what information-processing operations
a neuron carries out depend upon the way(s) its output is interpreted by
other parts of the system). This is why the detailed pattern
analysis of Abeles et al and Lestienne & Strehler is probably 
needed in addition to statistical approaches based on stationary processes.

It should be noted that average rates are one dimensional, scalar codes,
whilst temporal codes can support spike trains conveying more than one 
independent signal type (multiplexing). I think that this property of 
temporal codes has potentially very great implications for the design of 
artificial neural networks, if only because a multiplicity of orthogonal
signals allows one to keep different kinds of information from interfering
with each other.


Dr. Peter Cariani
Eaton Peabody Laboratory of Auditory Physiology
Massachusetts Eye & Ear Infirmary
243 Charles St., Boston, MA 02114 USA

email:  eplunix!peter at eddie.mit.edu
tel: (617) 573-4243
FAX: (617) 720-4408