One traditional use of Boltzmann machines is to perform pattern completion. In this task, the network is given degraded versions of the 8 digits from the digits task (each bit has been removed with 60% probability) and it must clean up the patterns to the full digit.
There are again two networks in this example: Boltz-Complete and Boltz-Complete2. The former has a single layer, the latter has two layers.
Let's begin with Boltz-Complete. Notice that it has a single layer that is both an INPUT and an OUTPUT group. The positive phase is not very interesting in this network. Units that have external inputs specified will be clamped by them. Otherwise the units are clamped to their target values. Since all units have targets, they are all clamped and the positive phase takes only one tick.
In the negative phase, only the units with inputs are clamped. The others are reset to the init value and are then free to update. Try training the network now.
Check out the network's performance in the Unit Viewer. You should find that it is mostly able to complete the patterns, but not always.
The Boltz-Complete2 network is similar, but it has distinct input and output groups. Train that one now. You'll probably find that its performance is a bit better than Boltz-Complete, but that it doesn't get as close to the targets because it takes longer to settle.
Choose one of the networks and see if you can get it to learn to complete all of the training set patterns. You may need to adjust the gain parameters, the learning parameters, or use weight decay.