I was thinking about training a model on non-linguistic material like video, and I was wondering if it could form concepts about the world, and also somehow form composite concepts or conceptual relationships, understanding that a “cat” is something that can “jump”, but non-verbally.
I was wondering if there is any unsupervised learning algorithm which is not designed in advance to look for any one kind of relationship over any other but could somehow be open to any and all sort of “patterns” in some data.
Like, I think a lot of unsupervised learning in linguistics is based on embeddings, but the choice of embedding can affect what the model ends up “identifying” - different embeddings can capture different aspects, different types of information.
Is there any algorithm which does not limit itself to look for one thing over another but somehow, mathematically, can sort of find correlations from a small scale up to higher and higher levels?
I was thinking that those “higher levels” would build on any concepts it had learned in previous stages. If it has gotten to the point where it resolves a cat’s face and body into eyes, red, arms, ears, it notices correlations between those and slowly builds a concept “cat”. Later, it may find correlations between “cat” and other things, like, “jump”.
Is this resemblant to a particular approach?