BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//CAIL//Events//EN
CALSCALE:GREGORIAN
METHOD:PUBLISH
X-WR-CALNAME:CAIL Events
BEGIN:VEVENT
UID:2026-10-30-stephen-tu@cail.columbia.edu
DTSTAMP:20261027T165002Z
DTSTART:20261030T150000Z
DTEND:20261030T160000Z
SUMMARY:ML Seminar: Stephen Tu - Multistep Backwards Latent Losses Learn U
 nstable Linear Dynamics
LOCATION:SSW 903\, School of Social Work\, 1255 Amsterdam Ave\, New York\,
  NY 10027
GEO:40.81022;-73.958347
DESCRIPTION:Model-based control synthesis relies on an accurate descriptio
 n of the system dynamics and the environment. When such models are unavail
 able or difficult to derive from first principles\, they must be learned f
 rom system trajectories. Latent world models have recently emerged as a pr
 omising general-purpose approach\, learning low-dimensional representation
 s together with dynamics that evolve these representations to predict futu
 re behavior. However\, it remains unclear which dynamical properties these
  models preserve\, and whether they support control synthesis with guarant
 ees of closed-loop stability and safety. Resolving these questions is esse
 ntial to establishing latent world modeling as a reliable tool for control
  design. In this talk\, we initiate a study of the dynamical properties ca
 ptured by latent world model designs. For trajectory data generated by a l
 inear dynamical system\, we show that multi-step backward latent reconstru
 ction losses recover the system’s unstable subspace\, provided the laten
 t dimension is large enough to represent the unstable modes. We then show 
 that an H-infinity controller synthesized using the learned latent dynamic
 s stabilizes the original system. Crucially\, both multi-step prediction a
 nd reconstruction in the original space (i.e.\, backward reconstruction) a
 re necessary for this guarantee: removing either can lead to an unstable c
 losed-loop system. Although our analysis focuses on linear systems\, exper
 iments on high-dimensional nonlinear systems reveal behavior consistent wi
 th our theoretical findings.\n\nhttps://cail.columbia.edu/events/2026-10-3
 0-stephen-tu
URL:https://cail.columbia.edu/events/2026-10-30-stephen-tu
END:VEVENT
END:VCALENDAR