Forgetful Memories
J. Leo van Hemmen, G. Keller, Reimer Kühn · Europhysics Letters (EPL) · 1988
Iterative learning schemes for fully connected neural nets are solved analytically. The key to the solution is a Markov chain representation of the nonlinear, iterative, encoding procedure, whose asymptotics can be determined exactly. Forgetting is an intrinsic property of the network, induced by a first-order transition of the retrieval quality as a function of the storage ancestry. The storage capacity, which can be obtained analytically, is extensive. Numerical simulations are found to be in excellent agreement with our results.