Theory of abstraction


I’ve been thinking about a theory of abstraction in the context of HTM systems. My question is on what research has been done into feeding htm systems into htm systems? I have some ideas that seem theoretically sound to me, but would like to know if there already exists some experimental or computational observations.


Are you talking about abstract compsci models?


No, mental abstractions in terms of sdr systems.


Does hierarchy fit what you are talking about? Their current focus is a single cortical region, although Numenta did some experiments with hierarchy before the current focus on objects and locations.

Do you mean something closer to intelligence than object recognition?


Abstraction is a hierarchy above object recognition, but I wouldn’t classify feeding and htm into an htm as an abstraction.


I think you’re talking about abstraction as in tying information to things that don’t have a literal space in the real world.

In literal object representation, we can tie the sensory features we’ve observed in any space to allocentric object space. For example if suddenly all coffee cups in the world had a new bumpy texture, over time your model of coffee cups would slowly change as you handled more and different cups with this texture in the real world.

I think about abstract concepts in the exact same way, except we create them, they are not observed directly by sensors. We use our experience with objects in space to imagine new objects that we have not observed in space. The first example we see of this in archeology is the lion man. The artist took two known concepts that had been observed in nature and created something new in an allocentric space. Then they literally created it in reality, because that is what artists do.