Learning to Transform Time Series with a Few Examples

Speaker:  Ali Rahimi

 I describe a semi-supervised regression algorithm that learns to
 transform one time series into another time series given examples of
 the transformation. I apply this algorithm to tracking, where one
 transforms a time series of observations from sensors to a time series
 describing the pose of a target. Instead of defining and implementing
 such transformations for each tracking task separately, I suggest
 learning a memoryless transformations of time series from a few
 example input-output mappings. Our algorithm searches for a smooth
 function that fits the training examples and, when applied to the
 input time series, produces a time series that evolves according to
 assumed dynamics. The learning procedure is fast and lends itself to
 a closed-form solution. I relate this algorithm and its unsupervised
 extension to nonlinear system identification and manifold learning
 techniques. I demonstrate it on the tasks of tracking RFID tags from
 signal strength measurements, recovering the pose of rigid objects,
 deformable bodies, and articulated bodies from video sequences, and
 tracking a target in a completely uncalibrated network of sensors.
 For these tasks, this algorithm requires significantly fewer examples
 compared to fully-supervised regression algorithms or semi-supervised
 learning algorithms that do not take the dynamics of the output time
 series into account.