In Jeff’s book he points out that in biological systems information is also sent down the hierarchy. Has this feature been implemented in any of your work? The idea seems intuitive. If I see a house I expect to see a door and a door knob. Potentiating the system to be sensitive to these features given the context makes sense.
I see lots of work on forward and backwards in time but have not noticed down the hierarchy but I am not an expert in HTM. This idea seems like it would be useful also in neural networks without a sense of time. But I have never seen it done.
As far as I know NuPic only support standalone regions.
Interconnectedness and the exact process of how it happens is still worked out, the theory of it anyhow.
If memory serves me well Jeff mentioned that they are heavily testing current ideas of ways to do that.
We’ve done a little bit of work on this topic in the research repository. In these experiments, feedback adds additional context for the temporal memory. This context can be used to disambiguate inputs when there is a lot of noise and help the system arrive at a solution quicker.
This directory contains the latest, including a two-page abstract with some results (see the PDF in there):
Hope this helps,