Zoom Meeting: https://umass-amherst.zoom.us/j/2614740715?pwd=UGprVTNjY2JSbHRodzZaZzRYcU1Qdz09
Meeting ID: 261 474 0715
Passcode: 12152020
Abstract:
Irregularly sampled time series data arise naturally in many application domains including biology, ecology, climate science, astronomy, geology, finance, and health. Such data present fundamental challenges to many classical models from machine learning and statistics. The first challenge with modeling such data is that the dimensionality of the inputs can be different for different data cases. This occurs naturally due to the fact that different data cases are likely to include different numbers of observations. The second challenge is the lack of alignment of observation time points across different dimensions in multivariate time series. These features of irregularly sampled time series data invalidate the assumption of a coherent fully-observed fixed-dimensional feature space that underlies many basic supervised and unsupervised learning models. However, there has been significant progress over the last decade on developing more specialized models and architectures for learning from irregularly sampled multivariate time series data within the machine learning community.
In this thesis proposal, we focus on the development of deep learning models for the problems of supervised and unsupervised learning from irregularly sampled time series data. We begin by introducing a computationally efficient architecture for whole time series classification and regression problems based on the use of a novel deterministic interpolation-based layer that acts as a bridge between multivariate irregularly sampled time series data instances and standard neural network layers that assume regularly-spaced or fixed-dimensional inputs. The architecture is based on the use of an RBF kernel-based interpolation network followed by the application of a prediction network. Next, we show how the use of fixed RBF kernel functions can be relaxed through the use of a novel attention-based continuous-time interpolation framework. We show that using attention to learn temporal similarity results in improvements over fixed RBF kernels and other recent approaches in terms of both supervised and unsupervised tasks. Finally, we propose to investigate the extension of these modeling approaches to temporally multiscale time series data including the problem of multi-scale imputation.
Committee Members: Benjamin Marlin (Chair), Madalina Fiterau Brostean, Dan Sheldon and Yan Liu (outside member)