Statistics Group



Bayesian inference and computation

Byrne, Chiappa, Cosma, Dawid, Gramacy, Graves, Lawrence, Parry, Spiegelhalter

Bayesian inference and computation provide a principled framework and a powerful toolbox for tackling challenging data analysis problems, such as detecting temperature changes over large timescales, or allowing robots to imitate human movement. We have

  • developed a principled approach for inferring the graphical model structure underlying observed data.

  • proposed a new stochastic computational algorithm to model long-term statistical dependence, allowing accurate modeling of complex environmental phenomena.

  • developed new methods for predicting future abundances of species in ecological statistics.

  • developed a novel probabilistic model for segmenting time-series such as human-movement data into basic actions for robot imitation learning. Information geometry treats a statistical model as a geometric structure. We have developed extensions based on `proper scoring rules', which motivate honest assessment of uncertainty.
   
   
SIFig barret
A statistical model as a geometric manifold. Top: robot arm used by a human for performing table tennis movements. Bottom: segmentation of a table tennis recording into basic movements as obtained by the model.

  1. Chiappa, S. and Peters, J. (2010) Movement Extraction by Detecting Dynamics Switches and Repetitions} Advances in Neural Information Processing Systems, 23, 388--396.

  2. Dawid, A. P. (2007) The geometry of proper scoring rules Annals of the Institute of Statistical Mathematics, 59, 77--93.

  3. D. Barber, A. T. Cemgil, and S. Chiappa (Editors), 2011 Bayesian Time-Series Models, Cambridge University Press.