Seminar: Sampling meets machine learning

Abstract: Generating samples from an un-normalized probability distribution is a fundamental task in computational inference. Beyond Bayesian statistics, it has numerous applications in modern inverse problems such as image recovery, protein generation, or probabilistic inference in large language models. In this talk, we argue that Sequential Monte Carlo (SMC) methods, also known as particle filters, constitute a natural framework to tackle the sampling problem by exploiting the parallelism offered by modern hardware (such as GPU). We point out two main practical challenges of SMC, namely sample waste and path degeneracy, and how to get around them. Finally, we highlight the flexibility of the SMC framework and incorporate latest advances from the literature on diffusion models , well-known for their image generation ability, into the sampling context. All in all, this shows the potential of SMC to combine techniques from the mathematical statistics and the machine learning communities to deliver more efficient inference algorithms.

Bio: Dr Hai-Dang Dau is a postdoctoral research fellow at National University of Singapore. His research interest is computational statistics, with a particular emphasis on the interaction of Bayesian computation and modern generative modelling techniques such as diffusion models and transport flows. He holds a PhD degree from Institut Polytechnique de Paris under the direction of Nicolas Chopin and was a member of Arnaud Doucet's research group at University of Oxford, where he applied his expertise in Sequential Monte Carlo methods to tackle the sampling problem.