[Ben.F.Rayfield]

http://en.wikipedia.org/wiki/Online_machine_learning

Online machine learning is used in the case where the data becomes available in a sequential fashion, in order to determine a mapping from the dataset to the corresponding labels. The key difference between online learning and batch learning (or "offline" learning) techniques, is that in online learning the mapping is updated after the arrival of every new datapoint in a scalable fashion, whereas batch techniques are used when one has access to the entire training dataset at once. Online learning could be used in the case of a process occurring in time, for example the value of a stock given its history and other external factors, in which case the mapping updates as time goes on and we get more and more samples.

The Contrastive Divergence algorithm, with annealing the temperature var down in each cycle, learning all the data again in each cycle, with the simplest form of dropout (weighted coin flip observe each bit var)... It works for 1 big batch, but it becomes unbalanced as more data is learned into it without relearning the old data.

If I just want to learn a stream of bit vectors and at any time be able to predict which bit vector (or some combination or permutation of them) any given input is most similar to, what are the algorithms normally used for that, or paths of research to explore?

[ranch vermin]

Thats the basic function,  to retrieve the closest oldest state.  You hop around alot checking the overlap of the features to get that.

But a boltzmann does something more than that,  very much related to it, what ive just worked out,  and its in the shape of the collections/features of the vector,  that this shape is an intercausation or an attention - (its looking at this pattern at this time.)  detected with the contrastive divergance       

Everything in the cell.  (which becomes a cause, when you look at it this way.)  is a way of looking at the situation, but it is just a piece of the vector the same...  If I brought all the causes back together, I can then rely apon them as whole situations that make sense together.  (like a tennis player grouped to the other tennis player, because they are relying on each other for their movement... which contrastive divergence would detect.)   

So then I can take a hard vector record, (like a movie) and by using contrastive divergence, to chip away at the video coming into it, I now have 'pieces of the action' which have reducted communication to the rest of the state/frame/vector  to now act independantly of what wasnt causing it!!!

[Ben.F.Rayfield]

The most basic function of a boltzmann machine is to statistical compress and from any vector reproduce the most similar vector it was trained on.

The second most basic function, which is far more useful, is to find permutations of permutations of combinations of parts of a literal data to a depth at most the number of layers, which can take the form of moving the data in space or many other transforms of it.

You have given so vague or abstract of analogies that I dont see how they relate to boltzmann machines. Do you have any math or algorithms?

[ranch vermin]

I dont understand Boltzmann machines fully, to go teach contrastive divergence,  its just that they all machine learning does a similar thing. 

It all boils down to something course and simple tho.
I read this guy talking about deep learning, and boltzmann machines were outperformed by much simpler algorythms - at making 'features'  and Im telling you can think about 'features' in another use of them,  as intercausal attention groups. (that was I was talking about - something original.)

Doesnt mean I hate boltzmann machines, and the ideas behind them help you immensely in other areas, but they are slower than other methods,  for example translational invarience better be using a pretty cut down restricted boltzmann machine only -  or youll be waiting so long because it makes an epoch from just for one frame of an epoch.

You might not know how computational a trans invarient net is, making boltzmann's not look so good.