IFT 3710/6759 - Active learning and Bayesian optimisation

Brief description

This project is about the implementation and exploration of active learning and Bayesian optimisation methods. Active learning refers to a class of machine learning method whose goal is to efficiently select which data points should be annotated next so as to maximise the information about an unknow oracle. The goal of Bayesian optimisation is similar, but its focus is on optimisation, that is finding the optimum of an unknown function. Both areas are of great interest for scientific discovery and experimental design. The goal of this project is to implement and study these algorithms with either synthetic data or highly controlled scenarios.

Data

This project is quite flexible and the students are welcome to propose any data to experiment with. However, we propose to first study the algorithms with synthetic data then extending towards benchmarks with practical application.

Expectations

As a flexible project, there is no strict expectations about the outcomes of the project, but the students are expected to provide an analysis of active learning or Bayesian optimisation methods under different conditions. For example, a project could focus on the sample efficiency of active learning algorithms using different acquisition functions.

References

• Paper: A tutorial on Bayesian Optimization of expensive cost functions, with application to active user modeling and hierarchical reinforcement learning
• Paper: Active Learning Literature Survey
• BoTorch tutorials