[Research] Reminder: Practice Job Talk Today 3:30pm NSH 3305

[Sajid M. Siddiqi]

Hi all,

Please come and share your feedback with me for my practice talk today, in
NSH 3305 at 3:30pm. Your input will be greatly appreciated. Thanks!

Title: Learning Latent Variable Models of Dynamical Systems

Abstract:
Dynamical systems and time series data are ubiquitous in our world. Models
for dynamical systems find widespread application for understanding these
systems, predicting their future behavior, time series classification and
recognition, planning and control, and much more. In this talk I will give
an overview of a popular class of dynamical system models, namely latent
variable models, with an emphasis on Hidden Markov Models (HMMs) and
Linear Dynamical Systems (LDS). I will describe work we have done on novel
learning algorithms for HMMs and LDSs that are more efficient, accurate
and robust than their predecessors in many ways. In particular, I will
focus on the problem of learning models of appropriate dimensionality, on
learning stable dynamics, and on learning the underlying manifold of a
dynamical system. I also briefly describe recent work on a hybrid latent
variable model that captures aspects of both HMMs and LDSs. I will present
results on human motion capture data for graphics, dynamic texture videos
for computer vision, and mobile robot camera data for robotics
applications.

Sajid