Architecture, constraints, and behavior


#1

I wasn’t able to find any existing discussion about this, but I’d be interested to get a sense of what the community’s thoughts on the material presented in this paper was.

In particular, there’s a nice analogy about textiles in how objects in the real world are generally found to be constructed or composed: there being several layers of fairly high diversity, with constraining protocols, or “glue code”, for organizing the diversity of a lower layer to that of a higher layer (e.g. fabrics -> seams -> clothes). The other major insight I had gotten from the paper is how that although there is incredible diversity within a layer (e.g. words in a language), when compared to what it could be (e.g. all possible combinations of sounds/letters), it’s actually rather small. So as you go higher up in the organizational hierarchy, object-space becomes larger and sparser.

How would something like this fit in with the more recent ideas in HTM theory about object compositionality (associating locations through displacement vectors)? Or how objects are represented throughout the hierarchy when there is so much “empty” space (though I suppose the answer to that is probably SDRs)?