Edwin V. Bonilla

Information for Potential PhD Students

If you are interested in doing a PhD in probabilistic machine learning at The University of New South Wales (UNSW) under my supervision please first read the following info:

If you have the necessary qualifications and background, feel free to contact me via email attaching the following documents in pdf format (I will disregard documents in any other format):

  • A CV.
  • A copy of the results of the UNSW HDR self-assessment tool described above.
  • A copy of your academic transcripts.
  • A 2-4 page research proposal describing the research you would like to do. Please see the list of potential PhD topics below.

You must have a strong math background and good programming skills. I will support your application as long as (a) you are indeed an outstanding student and (b) your project proposal  is aligned with my research interest

Potential PhD Topics

Your PhD proposal must be related to things I am actually interested in. As you may have seen on this site, I work in the general area of machine learning and am particularly excited by problems in probabilistic modelling and inference. Here a suggested list of topics/areas for your proposal:

  • Network structure discovery and causality. I always refer to the problem of determining the causes of a phenomenon (call it cancer, un- employment, depression, etc.) as the 'holy grail' of science and ultimately is the question that most scientists want to answer. One of the most difficult problems in this area is to determine what we can infer in terms of causal relationships from purely observational studies. See this paper and references therein as an example.
  • Automated probabilistic reasoning and implicit models. One of my long-term goals in probabilistic modelling and reasoning is to have flexible models and automated yet scalable inference methods for them. This covers the areas of implicit models and 'black-box' inference. See this paper, also this one and references therein. I believe that by having flexible priors, flexible likelihood models and flexible approximate posteriors, along with fast and scalable inference algorithms, we have the potential the revolutionise scientific research and Bayesian inference in general.
  • Structured prediction. Structured prediction is concerned with predicting multiple interdependent variables. My goal here is to develop the fundamental science for the next generation of structured-prediction algorithms, which will have a significant impact on a large variety of applications, such as social graphs, natural language processing, action recognition and computational biology. See this paper and references therein.
  • Bayesian deep learning. Deep learning has revolutionised application areas such as computer vision but there is strong evidence that these machine-learning frameworks are 'data-hungry' and can be very poor at quantifying predictive uncertainty (see e.g. this paper as a recent reference). My goal here is to develop sound uncertainty-quantification approaches in deep learning architectures. Our ICML 2017 paper with collaborators at EURECOM describes a step on this direction.
  • Applications. I am very interested in applications in the areas of renewable energy (e.g. solar energy forecasting) and climate change, natural language processing, machine learning for good (think health, fairness, etc), Bayesian optimisation and reinforcement learning.

The above list is by no means exhaustive and you are welcome to propose your own ideas, taking into consideration my research and publications. Please note that I am no longer willing to supervise students wanting to work on compiler optimization and that under no circumstances I will supervise a student willing to work on genetic algorithhms, evolutionary computation or any sexually-motivated computer science area.