A picture of the learning procedure

1. The bottom-up input causes an initial state of the code units. 2. This initial state settles to a state that takes the code energies into account. 3. Then we ignore the bottom-up input and do a little more settling to get the corruption. To learn the feed-forward connections:      the settled state is the target output      the initial state is the actual output. To learn the lateral connections:      the energy of the settled state is lowered      the energy of the final state is raised.

input

E

code