I have to implement a single layer (no hidden units) Boltzmann Machine for a class. I have an implementation done, but I'm not sure it's working correctly.
I'll describe what I've done and hopefully somone can point out if I've made any mistakes.
The BM machine model is a set of binary units in which each unit is connected to every other unit via a weighted connection.
First, we train the machine on a training set of examples to determine the weights of the connections.
This learning phase can be divided into two parts - one is the empirical correlation between the units of the examples in the training set and one is the correlation between these units according to some probability model.
At the end of learning, the probability correlation is subtracted from the empirical correlation and the resulting correlation is the weights for our BM.
We can then set our neurons to an input and run the activation model on them to recieve an output.
I realise as I type this that my understanding of the topic is really very fuzzy. Here's a few questions :
Having trained the machine on a set of examples, what should be the output when providing it with one of those examples as input? Will you always recieve that example back or will you occasionally get one of the other examples or even nonsense patterns?
I've found several RBM implementations online. How does a Restricted Boltzmann Machine differ from one with no hidden units? I'd like to be able to edit a RBM into what I've described so that I can make sure my implementation works the way it should.
[link] [11 comments]