webinar register page

Webinar banner
Darren Wilkinson - Compositional approaches to scalable Bayesian computation
Typical implementations of models and algorithms for Bayesian computation lack scalability, from a variety of different perspectives. Ultimately, scalability requires compositionality, but traditional approaches and programming languages are poorly suited to this way of thinking. Functional programming languages have seen increasing adoption in recent years, driven in part by big tech companies with a need to process data at scale. Expressing algorithms in a functional way is not only more elegant, concise and less error-prone, but provides numerous more tangible scalability benefits, such as automatic parallelisation and distribution of computation, as well as increased amenability to automatic differentiation. The Scala programming language is routinely used for building scalable data processing infrastructure, often in conjunction with Apache Spark for distributed big data processing and ML. Modern auto-diff frameworks such as JAX require a functional approach to algorithm development, and libraries such as BlackJAX show that they form a solid foundation for the development of gradient-based MCMC samplers. Experimental functional array languages such as DEX give an idea of what the future of ML programming languages may look like. Functional programming is intrinsically more compositional than traditional imperative programming. Strongly typed compiled functional programming languages are ideally suited to the development of scalable Bayesian modelling languages and computational algorithms, and this talk will attempt to justify this claim.

May 17, 2022 05:00 PM in Paris

Webinar logo
* Required information


Darren Wilkinson
Professor of Statistics @Durham University
Darren Wilkinson is Professor of Statistics at Durham University, UK. His current research interests involve applications of Bayesian statistics to a variety of challenging big data problems in molecular biology and engineering, and understanding the extent to which category theory and pure functional programming can provide a foundation for more scalable approaches Bayesian modelling and computation. He is especially interested in parameter inference for dynamic models, on-line inference for high-velocity time series data, probabilistic programming, and the use of approximate models and emulators for rendering computationally prohibitive algorithms for expensive models more tractable.