Hopfield discovered that his neural networks reproduced many of the properties of dreams identified long ago by psychologists, who found that we need to sleep and dream aftera series of exhausting experiences. he found that if he filled a neural net with too many memories (i.e., valleys), then the system began to malfunction from overload--i.e., the amount of time it took to access different memories began to become increasingly unequal. It began to malfunction in recalling previously learned memories. In fact, unwanted ripples began to form on the surface of the terrain that did not correspond to any real memories at all. These ripples are called "spurious memories" and correspond to dreams. Unlike real valleys, they do not represent real events, but are composed of fragments of existing memories.
In order to eliminate these spurious memories, he would add a small disturbance to the system, abruptly changing the terrain (so thee ball would be thrown out of a valley and would roll once again). The system was then allowed to settle down again into a state of deep energy minimization. Hopfield says this corresponds to sleep.
After several episodes of dreaming and sleeping, the system "awakened" refreshed--i.e., it stopped malfunctioning and could ercall all its memories at the same rate.
If Hopfield is right, then perhaps all highly developed neural nets, mechanical or organic, must dream in order to process their memories. Whenever a neural net is overloaded, it necessarily begins to act abnormally, creating memories that are not real--i.e., dreams consisting of random fragments of real memories. The system sleeps to cleanse itself of these fake ripples or dreams.
