MEET THE LECTURER: DR BRENDAN VAN ROOYEN
Queensland University of Technology / University of California, Berkeley
Born in South Africa and raised in Perth, Western Australia, Brendan has bounced around several great Australian and American institutions (UWA, UniMelb, The ANU and QUT/UC Berkeley). He now plies his trade as an ACEMS postdoctoral research fellow.
During working hours, Brendan is most likely to be found near QUT’s Y-Block or UC Berkeley’s Evans Hall. While tagged as a “statistician” or “machine learner”, he is broadly interested in all sorts of different mathematics, from category theory to concentration inequalities! In his day job, Brendan develops “new” scalable algorithms for solving machine learning problems and proves theorems regarding their performance. He is particularly interested in the problem of learning from data of variable quality and form.
In the usual theoretical analysis of machine learning algorithms, it is assumed that the data used to train our model of the real world, is of the same quality and form as the data we use to test our model’s conclusions. In practice this is rarely the case. Brendan’s research aims to understand and address this problem. In the process he has unhinged the support vector machine and created two machine learning frameworks, one in software, and the other in mathematics.
Outside of research, Brendan enjoys cooking, watching sports of all descriptions (particularly Rugby Union and Gridiron), and photography/digital art.
Dr Brendan Van Rooyen will be lecturing at AMSI Winter School 2017, delivering a course on “Martingales, McDiarmid and Machine Learning: How to validate models like a pro!”.
1. Can you tell us about your work? What drives your interest in this field?
I am primarily interested in learning problems where the “training set” is noisier than the “test set”. These problems appear all over the place in machine learning and statistics, where the “measurements” we can make about the world are not what would be like them to be. This may be because of a substandard measurement device (perhaps the expert cat labeller was sick….) or because of privacy/ethical constraints.
What keeps me interested in these problems is both the cool mathematics/ideas that go in to their solution, as well is the direct relevance to practical problems. While I spend my time proving theorems, it is nice to know I am close to practice!
2. What are the most interesting “big questions” or challenges facing researchers in your area?
Thanks to the explosion of online learning material as well as the growing availability of open source statistical software, almost anyone can apply machine learning to their problems. The biggest challenges therefore are not in practical deployment, but rather when one needs to do machine learning under various constraints, be they time, budget, measurement, fairness or privacy.
The big problems in machine learning therefore are all about understanding and assessing the impact of these constraints, be it creating new algorithms or assessing their statistical impact. Another key issue is simplicity. I would like to see Machine Learning become a well-founded engineering discipline. To do so one needs to identify the core of the field, the central base on which to build from.
3. What are some key industry applications of your work?
While Machine learning more broadly has found uses in countless many industries, my own work is currently confined to academic papers. While not directly applied (yet….), the process of writing these papers has allowed me to consult with various industries and provide them with high level direction, even if it is not specifically toward my work.
4. What do you consider your biggest achievement to date?
Surviving the first year of parenthood!
5. Why did you become a mathematician/statistician?
Personally, I do not like labels, I like problems. I began my university studies in Mathematics, Physics and Chemistry with a bit of programming thrown in.
I decided to focus on mathematics when I discovered that the core problems across these fields where much the same, although the mathematics exams were easier! Focusing on the key (albeit abstract) concepts freed up my time to do other thinking.
I became interested in statistics/machine learning during my masters. I saw it as a very thought provoking and eclectic field, with several competing sets of inferential principles (Bayesianism, Frequentism, pure and applied approaches to name just a few). I wanted to understand what was going on and this eventually led me to a PhD at The Australian National University working with Bob Williamson. Bob did (and does!) lots of mathematics although he is called an “engineer” or a “computer scientist”.
I am now a postdoc in the school of mathematics at QUT, however, the central goal of understanding and solving problems remains. Sometimes these problems are labelled computer science, sometimes statistics and sometimes mathematics, but that does not deter me.
6. Do you have any advice for future researchers?
Immerse yourself in your field. Read lots and gain a broad understanding of what the people around you are doing. Focus on the problems. Choose a research direction by finding out what irritates you in the current literature, by what part of the discussion you think is missing. Do not be afraid to ask questions nor to look stupid. Always hangout with people that are smarter/more experienced than you.
Basically put, if you are not confused, irritated and intimidated intellectually by the people around you, you are not doing research right! If this stuff was easy, it would not be research.