Advancing machine learning methodology for new classes of prediction problems.

  • Sanguinetti, Guido (Principal Investigator)

Project Details

Key findings

The research focused on methodology to improve learning by transferring knowledge across different machine learning tasks. As an initial subproject, we considered how to best improve classification by exploiting the availability of training data for multiple, similar tasks (multi-task learning). These results were published in the IEEE Trans on Neural Networks in 2011, and these ideas have been applied in a new international collaboration with ETH Zurich and TU Berlin to the problem of learning chemical properties of candidate drug compounds. Results from this very recent work have been published in early 2012 in the Journal of Computer Aided Modelling and Design. We then further extended the approach to incorporate information from unlabelled data (semi-supervised learning, paper under review). Finally, we considered the problem of extending knowledge from training on a task to a completely unknown task: this is supposed to be the most cognitively plausible mechanism of learning, whereby humans naturally generalise to previously unseen tasks. While of course we did not achieve human-like performance, the work represents one of the first attempts at this meta-generalising problem from a statistical angle, and is published in the top machine learning journal Journal of Machine Learning Research in early 2012.
Effective start/end date1/04/1031/03/11


  • EPSRC: £31,184.00


Explore the research topics touched on by this project. These labels are generated based on the underlying awards/grants. Together they form a unique fingerprint.