Forgetting as a way to improve neural‐net behavior

The output of neural nets of the Hopfield type favors some states more than others. Spurious states which are not in the learned set are also produced. It has been shown that by exciting the input of the net with random states and then ‘‘unlearning’’ the output state that the frequency of occurrence of the desired states becomes more uniform and that of spurious states is reduced. This has been proposed as one of the purposes of dreams.