I know the Hierarchical part of HTM isn’t developed yet, so I thought it’d be fun to think up ways of implementing it, even if those ways aren’t perfectly biologically representative.
Would it work to take the full or partial cell activation information of one or more temporal memory layers and input them into another one? From what I understand about the algorithms developed so far, that third layer would have the predicted futures at multiple points in time of multiple inputs as its own input. And, since there’s redundancy in whichever cells activate, only a few outputs would be necessary to predict average input from a large array of inputs.
Additionally, since the temporal memory algorithm has an inhibition radius, changing input from non-activating cells to other more active cells, like with a spatial pooler with locality added, seems like it would yield similar data to the original connection, but data that is more relevant at the time of the new input.
I know that’s likely not the way biology does it exactly, but it seems like it could yield some useful information.
Edit: Whoops, mixed up cells and segments. That’s embarrassing.
Oh, If I used small spatial and temporal memory layers for doing this compared to the ones reading the input, and read some of the input as well, would this be an SDR classifier?