5.5 Deep Boltzmann Machine
Deep Boltzmann Machines (DBMs) can be constructed from several RBMs where the hidden layer of the first RBM becomes the visible layer of the second, and so on, as shown in Figure 5.4.
Figure 5.4: Schematic representation of a DBM.
A DBM can be trained layer by layer, one RBM at a time. This will result in a powerful generative model capable of learning complex multivariate distributions and dependence structures. However, the generative training of the DBM can be used as the first step towards building a discriminative model if the training dataset samples are labelled. In this case, all DBM weights and biases found with the help of either CD or quantum Boltzmann sampling algorithms are seen as initial values of the weights and biases of the corresponding feedforward neural network. The discriminative model will consist of all the layers of the original DBM with an extra output layer performing the assignment of the class labels. The discriminative...
