[aaron]

Has anyone given any thought to how multiple HTMs will interoperate? For example, an HTM trained in subject A, working in conjunction with an HTM for subject B, in response to a query which needs (A+B).

Is there anyway of achieving dynamic on-the-fly integration of HTMs? Technically speaking 'merging' of Bayesian networks does not make any sense, but does it work to break down the query and then direct it to appropriate HTMs and aggregate the results?

That would be a neat thing to do if/when HTMs specialized in certain topics would need to talk to each other, e.g. over the web.

I think this is a very important topic that needs discussion and has not received as much attention. Any ideas?
AS

p.s. Perhaps there is a 'social' context to this question - would this be similar to how humans interoperate?

[FreeSynapse]

This is certainly a question that deserves a great amount of attention, especially by experts in the field rather than novices such as myself. The bigger question here is combining machine learning models, which really should be nothing new.

If the domains are the same, multiple model outputs can be weighted averaged. If the domains are different, the applicable model can be chosen. It also seems logical that in cooperative domains, one model would need another's outputs as part of its inputs, and vice versa, and this can possibly go on back and forth as needed.

As for strict merging of models, I guess the crude way to do it is to retrain over combinations of the underlying data that was previously used to train the individual models.

I wonder if there is any research into training machine learning models whose sole purpose is simply to combine existing models without retraining over their datasets. This may be one of the only generic ways of achieving on-the-fly integration of HTMs. That said, full integration of multiple models into a single one really doesn't seem to be crucial, because of the available alternatives of using domain specific models, weighted-averaging of multiple model outputs, or intermodel communication.

Further thoughts?