One remarkable feature of the resulting models is that they can develop very intuitive latent representations of their inputs. In these latent spaces, simple linear operations tend to give meaningful results,
