Abstract

Speaker 1 – Andrew Fitzgibbon

Title – Big data, small data, oddly-shaped data: Welcome to “All Data” AI

Abstract – I’m happy with the term “AI” — it just means doing cool stuff with data. We’ve seen great successes with computer vision, natural language processing, and a host of other applications. However, I’m not so happy when we shoehorn every problem into a BxWxHxC block of numbers to fit the constraints of GPU hardware. As the new head of the All Data AI (ADA) group at Microsoft Cambridge, I’m excited by a future where we can apply AI in traditional “big data” scenarios, in “small data” scenarios where we need to learn fast from limited examples, in the crossover area where we may have millions of related subproblems, each data-poor, but jointly data-rich. I’m excited to apply AI to structured data like graphs, molecules, program code. And I’ll talk about the compounding of excitement that results from applying these techniques to shipping products that impact millions of real users.

Speaker 2 – John Bronskill

Title – Fast and Flexible Multi-Task Classification Using Conditional Neural Adaptive Processes

Abstract – This talk will describe our recent work on designing image classification systems that, after an initial multi-task training phase, can automatically adapt to new tasks encountered at test time. I will introduce an approach that relates to existing approaches to meta-learning and so-called conditional neural processes, generalising them to the multi-task classification setting. The resulting approach, called Conditional Neural Adaptive Processes (CNAPS), comprises a classifier whose parameters are modulated by an adaptation network that takes the current task’s dataset as input. I will show that CNAPS achieves state-of-the-art results on the challenging Meta-Dataset few-shot learning benchmark indicating high-quality transfer-learning which is robust, avoiding both over-fitting in low-shot regimes and under-fitting in high-shot regimes. Timing experiments reveal that CNAPS is computationally efficient at test-time as it does not involve gradient based adaptation. Finally, I will show that trained models are immediately deployable to continual learning and active learning where they can outperform existing approaches that do not leverage transfer learning.