The design of experiments (DoE) is a problem that lies at the core of scientific discovery where the goal is to select an experiment (or a sequence of experiments) in order to achieve a specific objective, for example, (i) maximization of the information that the experiments provide about the underlying process; (ii) discrimination between two different models of a physical phenomenon and (iii) maximization of the design output. Application areas are diverse in, for example, psychology, pharmacology, physics and manufacturing. Although recent advances in machine learning have broadened the applicability of DoE methods, practical constraints in the real world have largely been ignored. These constraints include structured designs, where the input design is constrained to a feasible subset of the design space; multiple objectives, where one is concerned with designing experiments in order to address several objectives simultaneously (e.g. maximizing information gain and reaching an optimal outcome); non-stationarity, where the selection of a specific design may change the distribution of the outcomes; and efficiency and scalability, where the process of computing an optimal design is constrained by computational resources available or practical settings (e.g. online/interactive experiments).
This project aims at addressing the above problems by drawing from and developing new advances in deep learning, graph representational learning, probabilistic inference, Bayesian optimization and reinforcement learning along with Bayesian optimal design of experiments approaches. The outcomes of the project will have widespread applications in scientific discovery with and outwith CSIRO in areas such as synthetic biology and manufacturing.
The student is expected to develop the research and methods for the above problems, publish and present the corresponding outcomes at top machine learning venues (such as NeurIPS, ICML, ICLR, AISTATS) and contribute to specific applications involving DoE, e.g., in current projects at CSIRO (involving synthetic biology and/or manufacturing). The student will also be given the opportunity to visit our collaborators at Amazon (Berlin, world leaders in Bayesian optimization) and Max Planck Institute for Biological Cybernetics in Tübingen through internships and/or research visits.
- Computer science, statistics or related quantitative fields
- Excellent knowledge of machine learning techniques (e.g. popular supervised and unsupervised learning methods and fundamental concepts such as generalization, regularization, overfitting)
- Expertise in high-level programming languages such as Python
- Desirable: Knowledge of reinforcement learning, probabilistic inference techniques and deep learning frameworks such as Pytorch and TensorFlow