- 
                Notifications
    
You must be signed in to change notification settings  - Fork 8
 
Description
Add support to model hidden variables. Hidden units  should not interact with each other (no edges between them), for the ease of training. For Zephyr structured graphs there are four groups (colors) of qubits that can be used for this purpose.  Colors can be identified using `zephyr_four_color function of D-Wave network. The number of hidden variables that one can model depends on the number of qubits of one of the colors. For example, for zephyr Z6 there are approximately 300 qubits per color.
Addition of hidden variables complicates training. I suggest to marginalize hidden variables on the positive side of training.
See page 11 of this paper for calculation of free energy  :
https://www.cs.toronto.edu/~jmartens/docs/RBM_%20Representational_%20Efficiency.pdf