Me too. I think this is generally called "discovery" -- when a new concept arises because known concepts can't adequately explain an observation; established m

Me too. I think this is generally called "discovery" -- when a new concept arises because known concepts can't adequately explain an observation; established models can't capture it. IMO, it's easily the most interesting kind of learning. But frankly, I have no idea how a computational approach like machine learning could discover something new that's not just a combinatoric variation on known features. If you must invent a new dimension to model the unknown feature, how do you go about proposing one that's plausible and not absurd (like 4-dimensional space or 2-dimensional time)?