Numerical simulations of Boltzmann Machines
David G. Bounds · AIP conference proceedings · 1986
Statistical mechanics methods have been used to investigate the Boltzmann Machine algorithm proposed by Hinton and Sejnowski. Exact calculations of the partition function for a ten‐unit Boltzmann Machine show that there is a ‘‘window’’ of annealing temperatures at which learning is possible. For the 4−2−4 encoder problem it is found that the optimum learning rate is obtained when the number of energy states thermally accessible is approximately twice the number needed to store the hidden unit codes.