A New Variable Binding Mechanism
- Variable Binding is needed by neural nets when properties are
bound together for a short time.
- Imagine a red square and blue circle being presented to the system.
- Somehow the system has to tell the circle is blue even when both
red and blue are active.
- It's plausible that there is a CA for red square, but in many
instances there can be a combinatorial explosion CAs (big red
square with yellow polka dots).
- The standard mechanism is to bind via temporal synchrony.
Neurons in the bound CAs fire together. So red and square fire
in steps 1, 3, 7, and 10, while blue and circle fire in 2, 4, 6 and
9.
- We used a mechanism in our counting paper (cognitive modelling
conference last year) with our standard LTP Hebbian rule, but
bindings were erased by spontaneous activation.
- I've come up with a new mechism based on STP.
- The weights increase when neurons co-fire, but the weights
spontaneously decay.
- So, the binding is erased over relatively brief times, and can
be reused.
