Assessing the performance of training three stacked RBMs: Here, we will run five epochs (or iterations) for each RBM. Each epoch will perform batchwise (size = 100) optimization. In each batch, CD is computed and, accordingly, weights and biases are updated.

To keep a track of optimization, the MSE is calculated after every batch of 10,000 rows. The following image shows the declining trend of mean squared reconstruction errors computed for 30 batches for three RBMs separately: