Such a discovery furnishes a compressed, and there- fore more efficient, account of the data at hand. Thus, while scaling up data spaces, one may have to scale down the number of latent states generating those data, to the extent that doing so does not sacrifice accuracy.
there is a description of the data in terms of its gen- erating causes (e.g., the world of objects readymade) that is lower-dimensional, more predictable, and less ambiguous.