A Mean Field Theory Learning Algorithm for N e u r al N e t wo r ks
Carsten Peterson, James R. Anderson · Lund University Publications (Lund University) · 1987
Based on t he Boltzmann Machine concept, we derive alear ning algorithm in which time-consuming stochastic measurementsof correlations a re replaced by solutions to dete rminist ic mean fieldtheory equ ations. T he method is applied to t he XOR (exclusive-or ),encoder, and line sym metry problems with substantial success. Weobserve speedup facto rs ranging from 10 to 30 for these ap plicat ionsand a significan tly bet ter learning performan ce in general.