A Constraint Generation Approach to Learning Stable Linear Dynamical Systems Sajid M. Siddiqi, Byron Boots, Geoffrey J. Gordon (poster ID W22) · Convex Programming with Constraint Generation for learning stable Linear Dynamical Systems (LDS), applied to learning Dynamic Textures from videos

· A: Original training videos, smoke and fountain B: Latent states simulated from LDS models learned naively (least squares), an alternative stable learning method (Lacy Bernstein), and our method C: Simulated videos from least squares model (above) and our method (below)