Date: January 27th, 2021
Time: 10:00-10:45 IST
Lecture 1:
Identifying the Intrinsic Dimension of High-Dimensional Data
Abstract: One of the challenges of high-dimensional data is identifying the true information contained therein. In this presentation, I will describe some approaches that we have developed to address this challenge. These approaches include Bayesian methods of matrix factorisation, intrinsic dimension and structured variable selection. This discussion will be set in the context of substantive case studies in image analysis, sport and genomics.
Slides
Date: January 27th, 2021
Time: 11:00-11:45 IST
Lecture 2:
Finding Patterns in Highly Structured Spatio-Temporal Data
Abstract: Many forms of current data are indexed by space and/or time. Often these space-time relationships are highly structured. Examples include data collected across networks of streams, multi-state data collected over time, and multivariate responses collected at small area level across a geographic space. In this presentation, I will describe some of our approaches to Bayesian modelling and analysis of these types of data. Two particular issues will be discussed. The first is the role of priors in spatial models and hidden Markov models. The second is the detection of anomalies for event identification and to increase the trustworthiness of the data.
Slides
Date: January 29th, 2021
Time: 10:00-10:45 IST
Lecture 3:
Describing Systems of Data
Abstract: A common challenge is to analyse data as part of a system. In this presentation, I describe a number of Bayesian approaches to modelling such data, using as motivation case studies in neurology, ecology and industry. These case studies focus on understanding changes in the brain associated with a degenerative disease, and suggesting optimal dredgingstrategies for conservation of seagrass. The techniques employed include Bayesian wombling and dynamic Bayesian networks.
SlidesAbstract: Evidence-based decisions depend critically on trustworthy data. Two forms of data that have been brought into question in recent times are sensor data and citizen science data. Sensors are a key component of IoT and have created a step-change in our ability to monitor systems. However, they are often subject to technical anomalies that raise concerns about the validity of their data and signals. Citizen science is also growing in utility and interest in many areas, but often suffers from concerns about the credibility the information provided by community members. In this presentation, I will describe some new approaches to resolving some of these concerns. These include new methods for anomaly detection in high-dimensional streaming time series, and Bayesian models for estimating the latent ability of citizens taking into account the difficulty of the tasks. This work has been developed in collaboration with a number of teams working on challenges in ecology and industry; these teams will be acknowledged and the associated challenges discussed during the presentation.
Slides