In this example, a deterministic Boltzmann machine is trained to solve the simple digits task. You should begin by reading the manual page on DBMS.
There are actually two networks in this example, Boltz-small and Boltz-big. The former just has an input and an output layer, the latter adds a hidden layer. Let's start by training the Boltz-small network. Go ahead and hit train once or twice.
As mentioned in the manual page, a Boltzmann machine goes through two phases while training on an example, a positive and a negative phase. Click on an example in the Unit Viewer and step through the ticks in the example.
On the first tick all of the units are clamped to either their externalInput or their target value. The positive phase only lasts one tick in this network because there are no free units to settle. The next tick starts the negative phase. The output units have been reset to their initOutput, which is 0.5.
Now step forward to the end of the example. You should see the output units ramp up towards their target values. Use the Procedure pulldown menu to change to the Test procedure. Now when you click on a unit, only the negative phase will be run.
Now switch to the Boltz-big network and train it. This network has more trouble reaching the targets because it has two layers and it takes longer for activity to build up. So error will be higher. Set the Unit Viewer back to the Train procedure and click an example. Now you should see a distinct settling process in both the positive and negative phases.
Experiment with changing the network's initGain, finalGain, and annealTime parameters, which control the automatic adjustment of the network's gain during settling. The annealTime is the half-life, in time intervals, of the decay from the initGain to the finalGain. You may find that if the annealing is too rapid or the finalGain is too high, the network becomes unstable. But if the gain remains too low, it will be hard for the network to drive the units towards the targets. To train a Boltzmann machine effectively, it's important to find a good balance.