Variable Binding
- A classic criticism of neural nets is that they
can't do variable binding.
- The classic mechanism (Van Der Marlsberg 1986) to resolve this
is with synchronous firing.
- VB is a host of problems that are lumped together but one
classic problem is the green square and the red circle.
- All 4 base CAs are active, but how can you tell the colour of
the circle?
- Somehow they have to be bound together pairwise.
- Sougne has solved it using pattern oscillations, but
I think this has some real problems.
- I've solved it using change in synaptic weights.
- There are base areas for colour and shape, and a binding area
of medium term memory.
- The binding node can bind red and circle together.
- In the simulations, there are 2 nets with 10 base CAs each.
- The binding area has some different constants. (Spontaneous
activation 3% vs. 0, saturation base 24 vs. 21, threshold 7 vs. 5,
and decay 5 vs 1.5. )
- The idea is that the binding area both learns faster and
forgets faster.
- After one presentation of a pair, it is bound.
- Spontaneous activation wipes out the binding after 1200 cycles.
- This binds correctly 90% of the time, with 5% retrieval of
unbound items.
- This is a different memory system than the
short and long-term memory system in semantic memory.
- It can then support the long term formation of a red-circle
CA (outside the binding area), or
- Variable binding also gives us rules, which makes the
whole system largely Turing complete.
- I've submitted this to the AAAI conference
