A sample from the data distribution

Single layer models:

A sample from the full TRBM
A sample from TRBM-VH
A sample from TRBM-VV
A sample from TRBM-HH


Two layer models (its better than the single layered results):

A sample from the full TRBM, 2nd layer
A sample from TRBM-VH, 2nd layer
A sample from TRBM-VV, 2nd layer
A sample from TRBM-HH, 2nd layer


Training:

Every model was trained for 10000 sequences of length 100, with learning rate of 0.0005 and momentum 0.9.

the above results were obtained using this implementation.

Additional Results: this is what we get when we train each layer for 20K iterations rather than for 10K iterations. We also trained a third hidden layer. To get these results, modify every line word of the code that contains "10000" with "20000". There are only four such modifications that need to be made.
Single layer models:

A sample from the full TRBM
A sample from TRBM-VH
A sample from TRBM-VV
A sample from TRBM-HH


Two layer models (its better than the single layered results):

A sample from the full TRBM, 2nd layer
A sample from TRBM-VH, 2nd layer
A sample from TRBM-VV, 2nd layer
A sample from TRBM-HH, 2nd layer


Three (its better than the single layered results):

A sample from the full TRBM, 3rd layer
A sample from TRBM-VH, 3rd layer
A sample from TRBM-VV, 3rd layer
A sample from TRBM-HH, 3rd layer