@ylecun
@YiMaTweets Would you have a standard reference for *training* a system from data (system identification) with sufficient flexibility to be trained with a "straightening" criterion? Obviously, using a locally-linear approximation of a non-linear system is standard practice. But what we're doing is different: we are training an encoder (that maps observations to states) so that state dynamics follows trajectories with minimum curvature. The basic idea is not new. This was the topic of Olivier Hénaff with @EeroSimoncelli at NYU.