2020 Award Recipients
ETH Zürich, Switzerland
Thesis title: Privacy-Preserving Data Analytics at Scale
Recent years have seen unprecedented growth in networked devices and services that collect increasingly detailed information about individuals. With this trend comes the problem of ensuring the privacy of user data. To resolve this issue, many recent research efforts are exploring how to build data processing systems that follow the end-to-end encryption paradigm, where data is encrypted at the source such that services never see data in the clear. Existing encrypted data processing systems show great promise by allowing for confidential computation, but they are often limited to a few aspects of the system design. Important functionalities such as notions of data ownership, selective release of information, or even guarantees about the robustness of the computations are missing. In this thesis, we aim to design a new class of encrypted data processing tools and systems that expand to the requirements of more complex applications. Hopefully, this work will facilitate more adoption of end-to-end encrypted systems.
University of Cambridge, UK
Thesis title: Continual Learning and Federated Learning with Neural Networks
Current machine learning systems excel at training on centrally available batch data. My research considers two scenarios that are more representative of many real-world applications. In continual learning, data arrives sequentially, and past data cannot be revisited (because of computational or privacy constraints). In federated learning, data is split among many different clients and cannot be shared (such as private data on mobile phones). Current approaches to handle these real-world constraints tend to be expensive and inaccurate. My research looks at designing new scalable algorithms for these two fields, exploring and taking advantage of the similar technical challenges faced in both. I will apply Bayesian deep learning methods to do so. These have the potential for good calibration for decision making and uncertainty propagation for continual learning. Previous methods almost exclusively focus on parameter uncertainty, but I will develop methods that represent function uncertainty, considering changes in the network output directly.
University of Oxford, UK
Thesis title: Rapid Adaptation of Artificial Agents
Deep Reinforcement Learning methods have led to exciting successes in many areas, such as robots learning to pack boxes or artificial agents playing (video) games at human level. Often, however, these agents are trained from scratch on a single domain, without the ability to adapt to environmental changes. In this thesis, we propose several theoretical and practical advances to enable agents to rapidly adapt to changes in their environment, such as new game levels or maps, or new players in a multi-agent game. The first part of this work focuses on enabling agents to maintain a belief about how the world works, do strategic exploration in unknown environments, and incorporate their uncertainty into the decision-making process. The second part focuses on skill adaptation when agents encounter new situations that cannot be solved with their existing skill-set alone. Our work allows agents to adapt faster and more efficiently than previously possible.