Figure 3

Average Log-Likelihood achieved on an \(n_v = 24\) dimensional shifting bar data set as a function of DBM shape, \(\alpha\). The total number of hidden nodes are kept fixed at \(n_{h^{(1)}} + n_{h^{(2)}} = 22\). An ensemble of 50 DBMs were trained with CD-1 and MA, where the solid line shows the median average log-likelihood achieved after \(10^5\) gradient iterations with a linearly decaying learning rate \(\epsilon = 1 \rightarrow 0.01\), and shaded regions delineate the 95th and 5th percentiles.