Comparison to Hopfield and Other Attractor Nets
- It turns out that this is really similar to Hopfield
Nets and Other Attractor Nets
- In particular they all settle into stable states.
- There has been a fair amount of study from Mean
Field Theory (Statistical Mechanics) of this type
of network.
- However we seem to have a proof
that is at odds with some basic work on Mean Field Theory
- According to Mean Field Theory, you should only be able to have
O(N) stable states where N is the number of neurons.
- Our proofs make it O(N^k) where k is any sized constant.
- We haven't worked it out but it seems the Mean Field Theory
proofs are based on about half of the neurons being on, or
on the patterns being random, but our proof assumes a
relatively specific type of stable state.
- I think this means that the work from the Physicists has
really missed the point of a complex calculation and
memory system.
- Moreover, a CA network can move on to new stable states, while
standard attractor nets cannot.
