Monday 30 January 2017, 2.00PM to 15:00
Speaker(s): Dr Charles Blundell, DeepMind
Recently deep learning methods have attained super-human performance on a wide range of tasks, from image classification to game playing. A common limitation of these methods is a requirement for a large amount of data from each new task.
Meta-learning provides a potential means of overcoming this limitation, where the general principles learnt on one task may be re-used to learn a new task more efficiently than starting from scratch.
I shall describe recent work we have done on learning to learn in the context of deep neural networks applied to image classification and reinforcement learning tasks.
Charles Blundell is a staff research scientist at DeepMind. He completed his PhD in machine learning at the Gatsby Unit, UCL, and before that studied computer science at York.
Please contact Helen Fagan, Postgraduate Office, for more information.
The Department also runs a programme of Research Student Seminars given by PhD students in their 3rd year of study.