Q&A with Chris Maddison
In 2019–20, Chris J. Maddison, Member in the School of Mathematics and a Senior Research Scientist at DeepMind, is developing methods for machine learning and exploring foundational questions about how learning from data is possible.
What drew you to this field initially?
Imagine you knew nothing about baking, but someone gave you a million different muffins. Could you figure out how to bake a muffin? That's the problem of machine learning.
Relatively few species learn the vocalizations that they produce. Among them are humans, dolphins, and songbirds. At the outset of my undergraduate studies I was interested in this question of how songbirds learn to sing. Towards the end of my studies, my interests had drifted into more foundational questions of how learning from data is possible in the first place. I became enamored with the algorithms of modern machine learning, and I've stuck with this field ever since.
Why IAS?
The success of deep learning is owed to academic groups that pursued their research directions driven by curiosity and independently of the whims of academic trends. I was looking for a place to pursue curiosity wherever it led, connect with like-minded researchers, and focus on research. IAS seemed like a great place to do that; the spirit of curiosity is at the core of the Institute's culture and embodied in the title of Abraham Flexner’s classic essay, “The Usefulness of Useless Knowledge.”
What question within your field do you most want to answer and why?
An exciting question in machine learning is how to use apparently unlabeled data to improve performance on multiple downstream tasks. This is sometimes called unsupervised learning. It’s a practical problem, because most data comes unlabeled by humans, but it is also a problem that is not paradigmatically settled. There are different proposed frameworks for unsupervised learning, but so far nothing emerges as a clear winner. There are even disagreements on how to measure progress!
Where is your favorite place to think?
When I am stuck in my research, a walk down a quiet woodland path is the best thing I can do to become unstuck.
What do you hope the impact of your research will be now or in the future?
Some of the best moments in the study of computer science are when we collectively realize something is possible that was once considered improbable. We’ve been very lucky in machine learning to have many such moments; among them are the unrelenting progress of image classification algorithms, AlphaGo’s win against Lee Sedol, and recent language models that can produce paragraphs of realistic artificial narratives. One of my goals is to contribute to these moments of surprise, to elicit the thought, “huh, that’s cool.”
How do you describe your work to friends and family?
Imagine you knew nothing about baking, but someone gave you a million different muffins. Could you figure out how to bake a muffin? That's the problem of machine learning.
What other activities or subject areas do you enjoy?
I enjoy reading poetry. I find it similar to the experience of reading a surprising proof or argument. When these are carefully written you can share in the moment of insight with the writer, and suddenly find yourself in a strange new world. I also enjoy making pottery on the wheel.
