Probabilistic Programming Workshop

Thursday 16 November (afternoon)

1. Introductions (13.00 - 13.15)
2. Session 1: 13.15 - 14.45

• Invited talk (Frank Wood (Oxford University) 40 min + 5 min questions)
• Invited talk (Sebastian Funk (London school of hygiene and tropical medicine) 40 min + 5 min questions)

3. Coffee break (30 min)
4. Session 2: 15.15 - 16.45

• Invited talk (Ewan Cameron (Oxford University) 40 min + 5 min questions)
• Invited talk (Hongseok Yang (Korea Advanced Institute of Science and Technology) 40 min + 5 min questions)

Friday 17 November (morning)

1. Session 3: 9.00 - 10.30

• Invited talk (Fredrik Ronquist (Swedish Museum of Natural History) 40 min + 5 min questions)
• Internal talk (Lawrence Murray (Uppsala University) 40 min + 5 min questions)

2. Coffee (30 min)

1. Frank Wood
2. Sebastian Funk
3. Hongseok Yang
4. Ewan Cameron
5. Fredrik Ronquist
6. Lawrence Murray

Frank Wood - Probabilistic Programming and Inference Compilation, or, How I Learned to Stop Worrying and Love Deep Networks

Probabilistic programming uses programming language techniques to make it easy to denote and perform inference in the kinds of probabilistic models that inform decision-making, accelerate scientific discovery, and underlie modern attacks on the problem of artificial intelligence.

Deep learning uses programming language techniques to automate supervised learning of program parameter values by gradient-based optimization.

What happens if we put them together?

This talk will review probabilistic programming. It will also introduce inference compilation and address how linking deep learning and probabilistic programming is leading to powerful new AI techniques while also opening up significant new research questions.

Sebastian Funk - Learning from the uncertain: modelling and forecasting of infectious disease outbreaks

Forecasting the course of an outbreak can inform public health planning and decision making. Accurate forecasts, however, are hampered by uncertainty about changes in human behaviour, pathogen genetics and environmental factors that are difficult to capture in real time. I will present how semi-mechanistic models can be used to make inferences in the presence of such uncertainties, and illustrate this approach with applications to recent outbreaks of Ebola, cholera and measles. Using these examples, I will discuss the methodological and computational challenges involved, as well as opportunities for integrating different data sources to understand and predict the behaviour of infectious disease outbreaks.

Ewan Cameron - Essential statistical modelling challenges in malaria risk stratification

Accurate and timely risk maps are a powerful resource for malaria control programs aiming to efficiently allocate interventions and to assess progress towards eradication of this burdensome disease. Issues of data quality and missingness arising from under-resourced health systems, as well as the complexities of the disease interaction between human and vector populations, pose a number of statistical challenges to this endeavour. In this talk I will describe the nature of these challenges---approximate inference for large geospatial models; learning human movement kernels and health facility catchments; performing data imputation under difficult patterns of missingness; calibrating simulations of health system use and outcomes---and will give my thoughts (and solicit audience feedback) regarding the potential of probabilistic programming approaches in each case.

Hongseok Yang - Formal semantics of probabilistic programming languages: Issues, results and opportunities

In the past two years, I and my colleagues have worked on developing so called denotational semantics of probabilistic programming languages, especially those that support expressive language features such as higher-order functions, continuous distributions and general recursion. Such semantics describe what probabilistic model each program in those languages denotes, serve as specifications for inference algorithms for the languages, and justify compiler optimisations for probabilistic programs or models. In this talk, I will describe what we have learnt so far, and explain how these lessons help improve the design and implementation of these probabilistic programming languages and their inference engines. This talk will be based on several joint projects with Yufei Cai, Zoubin Ghahramani, Bradley Gram-hansen, Chris Heunen, Ohad Kammar, Sean Moss, Klaus Ostermann, Adam Scibior, Sam Staton, Matthijs Vakar, Frank Wood, and Yuan Zhou.

Fredrik Ronquist - An interactive PPL for statistical phylogenetics

Statistical inference based on phylogenetic models — models built around evolutionary trees — is widely used throughout the life sciences today. The field is completely dominated by Bayesian MCMC methods, which were introduced about 20 years ago. The flexibility and computational efficiency of this approach have resulted in an explosive development of phylogenetic models. It has been quite challenging for computational biologists to keep up with the rapidly expanding model space, and the field is dominated today by a plethora of software packages, each dealing with a specific subset of models. There is a clear need for more generic approaches to model construction and inference. We have tried to address these challenges by developing Rev, a PPL for statistical phylogenetics based on probabilistic graphical models. Unlike most other PPLs, Rev is designed for use in an interactive computing environment, allowing users to build phylogenetic models step by step, and examine the model components as they go. I describe some of the challenges involved in developing an interactive PPL. I also discuss the potential and the limitations of probabilistic graphical models in phylogenetics.

Lawrence Murray - Delayed Sampling and Automatic Rao–Blackwellization of Probabilistic Programs

We introduce a dynamic mechanism for the solution of analytically-tractable substructure in probabilistic programs, reducing variance in Monte Carlo estimators. For inference with Sequential Monte Carlo, it yields improvements such as locally-optimal proposals, variable elimination, and Rao-Blackwellization. The approach is demonstrated with some pedagogical examples, and an epidemiological case study of an outbreak of dengue in Micronesia, using a new probabilistic programming language called Birch. Joint work with Daniel Lundén, Jan Kudlicka, David Broman and Thomas Schön.