Learning within bounds and dream sleep
Tamás Geszti, Ferenc Pázmándi · Journal of Physics A Mathematical and General · 1987
In a bounded-synapses version of Hopfield's model (1984) for neural networks the quasienergy of a given memory, which is approximately equal to the depth of the corresponding energy well is calculated exactly by treating the change of a synaptic strength on learning as a random walk within bounds. Attractors corresponding to stored memories are found to be considerably flattened before serious retrieval errors arise. This allows dream sleep to be interpreted as random recall and relearning of fresh strong memories, in order to stack them on top of weak incidental memory imprints of a day.