You have some vector pattern recogniser that can respond to say a few hundred input patterns. When it sees a known context in the input vector (I) it outputs a vector (A). Both that output (A) and the input data (I) are fed to another pattern recogniser (output B). For a particular value of A, the value of B can be decided depending on some other context information in the input. This is tree like splitting. You can keep going with another pattern recogniser that takes I and B as input.
That should work with the random projection neural nets I have and also with Numenta’s neural pattern recognisers as well. In my case I understand how I should go about training the system, I’m not too sure how you would do that with Numenta neurons.
Could you describe the problem you are trying to solve with this?
Data driven neural nets without back-propagation. The idea is to focus learning (more tree branches) in parts of the problem state space where that is needed . I’ll let you know how it turns out in a few days. I’m going to try with a recurrent version I thought of.
There are some disparate alternatives to back propagation scattered around in research papers. It would be worthwhile to collate them, if anyone ever had the time.
Interesting. Sounds like Growing Neural Gas, which I know some people here are working with.