Highly Scalable Autonomous Time Series Analysis

Most machine learning systems learn models that apply at a single point in time and take little or no account of the dynamic interactions that have led to the current state-of-affairs. This project will develop new autonomous machine learning algorithms that learn from time series data. These data capture the evolution of events that have led to a given state. This capacity to learn from such data is critical to many tasks. It is only possible to distinguish in satellite imagery whether a crop is wheat or barley by considering the evolution of images in each location. No single image in isolation provides sufficient discriminatory detail. When using measurements of railway track performance to assess remedial maintenance, only by considering the evolution of the track condition is it possible to determine whether simple ballast tamping will be effective or whether ballast replacement is required.

Machine learning from information collected at multiple time points is called time-series analysis. This project builds on a large body of previous research in this field to produce techniques that scale to orders of magnitude larger collections of time series than the current state-of-the-art and to much higher-dimensional time series, as is required by our two exemplar applications, satellite image analysis and railway track maintenance.