Variable Binding
- One of the big problems for neural nets is the binding (variable
binding, dynamic binding, moving brushes) problem.
- Symbolic guys don't notice it, because it is a fundamental operation
on Von Neuman computers.
- The short version is take a variable and stick a value in it.
- The red square blue circle problem is an example. If there is
a red square and a blue circle, then all four CAs are firing. Which
colour is the square?
- That's hard for neural (and most other connectionist) systems to do.
- The standard answer to this is binding by synchrony (Von der Malsburgh
1981). Roughly neurons that are bound tend to fire at the same time.
So red and square are firing in one pattern (say the odd cycles) and
the blue and circle neurons are firing in another (say every third
cycle).
- Another solution is binding by active links (van der Velde and de
Kamps 2006). Roughly, a circuit is set up that changes dynamically
and retains the binding while firing.
- I've come with a solution based on short-term potentiation. (The
paper is under review with Connection Science.) Roughly
the synaptic strength between bound items is increased, but
then fades away over time.
- I think these are some good answers to Jackendoff's (and Fodor's)
criticism.
