Events: detail
CSML seminar: When transfer works and single task learning doesn't
- Hosted by:
- UCL Department of Computing
- Speaker:
-
Andreas Maurer
- Starts:
- February 06, 2008 at 11:00 am
- Ends:
- February 06, 2008 at 12:00 pm
- Location:
- University College London, Malet Place Engineering Building, Room 6.12, Gower Street, London, WC1E 6BT United Kingdom
- Maps:
Description
I discuss circumstances, under which upper error bounds for transfer learning are upper bounded by lower bounds for single task learning with essentially arbitrary algorithms. I will briefly introduce tasks of half-space learning from the uniform distribution on the sphere and then present some lower bounds on generalization performance for single task learning. Then I discuss transfer learning in this context. If the tasks are sampled from an essentially low dimensional set of halfspaces, individual training samples are small and many tasks have been observed in the past, then the transfer learner provable outperforms any single task learner with high probability.
- Registration required:
- Yes
- Free:
- Yes
Additional information
Visitors from outside UCL please email in advance
For more information
- Contact person:
- CSML Webmaster
- Email:
- webmaster [ at ] csml.ucl.ac.uk
- Website:
- CSML seminar: When transfer works and single task learning doesn't
