ANC Workshop - Massimiliano Patacchiola

Speaker: Massimiliano Pattachiola

Title: Two Issues in Meta-Learning and How to Fix Them

Abstract:

A major bottleneck of modern deep learning systems is the need of large, labelled datasets which are expensive to gather and annotate. A recent line of work has tackled this problem by defining the meta-learning framework, a powerful approach that can be used to train a model in the low data regime. Meta-learning is promising but is still affected by (i) serious scalability issues and (ii) low performances under unbalanced data. Regarding the first point, I will provide an overview of our recent NeurIPS submission presenting a new task sampler that allows meta-learning methods to scale to large images and datasets. For the second point, I will discuss a large empirical study where we have compared meta-learners under different sources of imbalance, summarizing some useful take-away lessons.

Bio:

Massimiliano Patacchiola is a postdoctoral researcher at the University of Cambridge, working in the Machine Learning Group under the supervision of Richard Turner. His project is funded by EPSRC and Microsoft Research and it aims at building new machine learning methods for real-world applications that are efficient, flexible, and automated. Before joining Cambridge, Massimiliano was a postdoctoral researcher in the Bayesian and Neural Systems group at the University of Edinburgh with Amos Storkey, and an intern in the Camera Platform team at Snapchat.