Temporal Codes

[jamie@atlas.ex.ac.uk]

Peter Cariani (4/5/95) wrote:

>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.

I also think this is an important point.  In particular, the variable
binding problem is precisely the problem of keeping information about one
entity from interfering with information about another.  For example,
keeping the representation of a red square and a blue triangle from being
interpreted as a red triangle and a blue square.  Another reason for
expecting temporal codes to be used for representing such binding
information is that a neuron will generally react the same way to a given
input regardless of the absolute time at which the input occurs.  This
property can be used to argue that a temporal synchrony representation
(Cariani's temporal code, option 2) of variable bindings inherently implies
systematicity.  I know of no other code that can be argued to imply such
generalization across entities.

Of course, an argument for expecting variable binding information to be
represented in a temporal code is in no way an argument against expecting
other kinds of information to be represented in other codes.  These kinds of
computational considerations do, however, tell us something about what kind
of information to look for in what kind of code.


James Henderson
Department of Computer Science
University of Exeter
Exeter EX4 4PT, UK