Supervised
learning for the Boltzmann machine is run in two
phases, both of which use a sample input um. The first phase, sometimes
called the wake phase, involves Hebbian plasticity
between sample inputs and outputs. The dynamics of the Boltzmann
machine play no role during this phase. The second phase, called the sleep
phase, consists of the network “dreaming”(i.e.,
internally generating) v(um) in response to um based on the current weights W. Then, anti-Hebbian
learning based on um and v(um) from um. It is also possible to use
the mean-field method we discussed in chapter 7 to approximate the average over
the distribution P[v|um;W] in equation 8.65
Supervised learning for Boltzmann machine
Supervised learning for the Boltzmann machine is run in two phases, both of which use a sample input um. The first phase, sometimes called the wake phase, involves Hebbian plasticity between sample inputs and outputs. The dynamics of the Boltzmann machine play no role during this phase. The second phase, called the sleep phase, consists of the network “dreaming”(i.e., internally generating) v(um) in response to um based on the current weights W. Then, anti-Hebbian learning based on um and v(um) from um. It is also possible to use the mean-field method we discussed in chapter 7 to approximate the average over the distribution P[v|um;W] in equation 8.65