Table 1.

Algorithm description of train deep belief network.

Input: Data D = {x}, desired layers K and nodes number for each layer Ni
Output: The structure and learned initialization parameters of the DNN.
1. Learn parameters θ1 for the 1st layer RBM from data.
For k = 2:K
2. Initialize the k-th layer RBM by unroll the k-1th layer RBM to the kth layer, of which parameters $W_k=W_{k-1}^T$
3. Refine the parameters of kth layer RBM from data vectors generated from k-1th layer.
Return: Structure and parameters of the stacked RBMs.

Greedy training process for deep belief network