ECE PhD Proposal Review: Cooper Loughlin

PhD Proposal Review: Unsupervised Machine Learning Approaches to Sequential Data Analysis

Cooper Loughlin

Location: Remote

Abstract: Analysis of sequential data is central to many fields of science and engineering. Often, sequences are collections of observations made over time and space with little or no contextual information. The goal of analysis may be to evaluate relationships, identify unusual observations, or forecast future behavior based on historical data. Unsupervised modeling of sequences (e.g., time series) can illuminate the underlying structure of the data and enable analysis.

In this proposal, we discuss a statistical model for multivariate time series and an associated inference algorithm. We develop a preliminary model for a particularly challenging class of multivariate time series where the observations are counts (non-negative integers) that are nonuniformly sampled in time. We develop a state space model and inference algorithm based on Monte Carlo integration and Expectation-Maximization. This preliminary work highlights some key challenges still to be addressed. In particular, continuously variable sampling intervals, computational complexity of sampling, and long-term dependencies among observations are properties of real data that are not handled well by the preliminary model. Recent developments in unsupervised sequence modeling using deep learning techniques are introduced including variational auto-encoders, recurrent neural networks, and ordinary differential equation recurrent neural networks. We propose utilizing these deep learning techniques to improve the state of the art in sequential data analysis.