A DBN is a multilayer belief network where each layer is an RBM stacked against one another. Apart from the first and final layers of the DBN, each layer serves as both a hidden layer to the nodes before it, and as the input layer to the nodes that come after it:
Two layers in the DBN are connected by a matrix of weights. The top two layers of a DBN are undirected, which gives a symmetric connection between them, forming an associative memory. The lower two layers have directed connections from the layers above. The presence of direction converts associative memory into observed variables:
The two most significant properties of DBNs are as follows:
- A DBN learns top-down, generative weights via an efficient, layer by layer procedure. These weights determine how the variables in one layer depend on the layer above.
- Once training is complete, the values of the...