Hopfield nets have units in states 0 or 1 in which the (N) units are probably fully interconnected. They attempt to ``memorise'' n different network states - a network state is an N long binary vector. If these states are , where
we shall set the weights by writing
in the case , and . Now load the network with one of the states to be remembered, .
implying stability, subject to noise from the terms; we assume enough ``independence'' among the to suppose this noise may be disregarded.
Observe from 6 that we have symmetry in the weights, , and define the energy of the system at a given instant by
If we elect to change the value of processor , we can see that the change in energy due to this change is
State changes therefore cause E to decrease monotonically, and ``stability'' represents a local minimum of E.