A theme through much of my research is reducing the computational resources required for learning without compromising model quality. This is an increasingly critical problem as modern AI is attributed a comparable carbon footprint to the airline industry.
We have transformed the field of time series classification, achieving state-of-the-art accuracy with many orders of magnitude less computation for both learning and classification - see Scalable learning of time series classifiers.
We have developed methods for preconditioning logistic regression and neural network models using closed form initialisations based on Bayesian network classifiers. These both reduce reduce learning time and improve accuracy - see Combining Generative and Discriminative Learning.
We have developed methods for Scalable Graphical Modeling and Efficient Bayesian Network Classifiers.
We have developed statistically efficient methods for Learning Decision Trees from Streaming Data.
Publications
[bibtex file=publications.bib template=publist-efficient-ml format=my-format sort=year order=desc process_titles=0]