This grouping has contributed to the theoretical understanding of learning dynamics in Recurrent and Recursive Neural Networks and explained the power of intermittent search in finding high quality solutions to difficult assignment optimisation problems in Self-Organizing Neural Networks.
In data modelling, probabilistic models we have developed allow strictly principled analysis and visualisation of multivariate data, applicable to discrete sets and multiple temporal sequences. The use of novel hierarchical model construction, differential geometric methods and generative topographic modelling enhance the models' explanatory power, enabling predictions from very small samples on large realistic data sets. This work led to a PPARC-funded project with Astrophysics that enabled time-delay estimation in gravitational lensing superior to state-of-the-art results in astrophysics, and the discovery of a new class of galaxies.
The grouping contributed an influential integrating survey of ensemble learning and techniques for diversity creation.
In vision-related work we developed new learning algorithms for acquiring probabilistic graphical models of human actions and new techniques for efficient reinforcement learning.
Future research includes extending our work on data modelling, e.g. to give a principled account of how a representation influences learning from very small samples. Applied work will continue on new machine-learning techniques, stimulated by challenges arising from astrophysics, biosciences (e.g. proteomics) and functional brain imaging.