Inverting generative processes


George Papamakarios

In machine learning, data is typically postulated to have originated from a model via a generative process, and learning the model boils down to inverting that process. Given a generative process that we can simulate forward but not observe, the challenge is to use simulation results to learn something about its internal workings. A generic and scalable solution to this problem needs to combine ideas from approximate Bayesian inference, generative modelling and density estimation. In this project, we aim to contribute to the above areas with novel methods and techniques, so as to facilitate inverting generative processes for practical problems.


Supervisors: Iain Murray & John Winn