Speaker

Timothy Hospedales

Title

Meta-Learning Fine-Tuning Strategies

Abstract

The mainstream contemporary workflow for applied AI is based on fine-tuning large general purpose pre-trained models for specific applications. This raises the question of “how to conduct fine-tuning?” as a central practical and research issue in achieving highly performant systems. In this talk I will discuss our work on meta-learning fine-tuning strategies such as adapters, selective-update masks, and optimisers. I will also touch on meta-optimising fine-tuning strategies for different downstream criteria, such fairness, “safety” and calibration.

Bio

I am a Professor within IPAB in the School of Informatics at the University of Edinburgh, where I head the Machine Intelligence Research group; ELLIS fellow; and Head of Centre, Samsung AI Research Centre, Cambridge where I also direct the Machine Learning & Data Intelligence Programme.

Language

English