RRR's Deep networks use stacked layers of Restricted Boltzmann Machines (RBMs) to map class attributes to a target class. In this case, we map numerical CPU performance counters (e.g. FLOPs, Instructions completed, L2 Hit Rate, etc) to power consumed over an interval (300 ms). 

Each layer of the network is an RBM which is trained to reproduce its input values (unsupervised), with the final layer being trained to map the previous layer's inputs to the target power bin.