I don't see any reason why an ML algorithm wouldn't be able to extrapolate to novel ideas. Once it has learned some high dimensional subspace based on the examp

I don't see any reason why an ML algorithm wouldn't be able to extrapolate to novel ideas. Once it has learned some high dimensional subspace based on the examples given to it, I suppose it could search for areas outside that boundary which are in some way likely to be interesting.