Figure 4
From: Mode-assisted joint training of deep Boltzmann machines
Average log-likelihood on the MNIST data set achieved after 100 epochs with mode-assisted training (MA) compared to the recent best achieved results on DBMs using CD as well as pre-training. DBMs trained with MA used PCD-1 for all non-modal samples, matching Ref.18. Network topology was fixed to \(\alpha = 0.15\) with an increasing number of hidden nodes, \(n_h\). The learning rate was chosen to follow a linear decay, \(\epsilon = 0.05 \rightarrow 0.0005\) and no pre-training was employed. The best log-likelihood was reported out of 10 randomly initialized runs.
