Simulation Based Inference in the Natural Sciences – Workshop

When: Friday 31^st March, 13:30-17:00 (drinks reception 17:00-18:00)

Where: Blackett Lab, Lecture Theatre 2

• How do you perform statistical inference in settings where the likelihood is intractable?

• Could your research benefit from hearing more about simulation based inference?

The FoNS Data Science theme invites researchers from across the faculty to a half-day workshop on the use of simulation based inference in the natural sciences. Featuring talks from researchers across FoNS, we will discuss modern, machine learning based approaches to SBI, as well as more established techniques, such as approximate Bayesian computation.

We encourage participants from all FoNS departments and across College , please register by 30^th March 5:00pm!

Note: we strongly encourage participants to attend this event in person, but if this is not possible you can join via Teams meeting.

Schedule:

• 13:30-13:35  Welcome from Professor Alan Heavens
• 13:35-15:00  Talks – Session 1
• 15:00-15:20  Coffee break
• 15:20-17:00  Talks – Session 2
• 17:00-18:00  Drinks & nibbles reception

Please read below full details about the speakers and their talks:

Session 1

Alan Heavens – Simulation-based inference for the uninitiated

An introduction for those who are new to SBI and interested in how it might be useful for their research. I will cover how SBI works, how it can be used in a principled Bayesian analysis, and how the extreme data compression that is often required can be achieved without losing much information (none in ideal cases), by using mathematical techniques, or with neural networks.  Finally I’ll touch on techniques of Approximate Bayesian Computation (ABC) and various flavours of neural density estimation which can be used to learn all of the important probability distributions needed for inference: posteriors, likelihoods, and sampling distributions.

James Martin – Approximate Bayesian computation: state space models and beyond

 In this talk, I’ll give an introduction to using ABC in computational inference, drawing attention to important statistical considerations. I’ll pay particular attention to performing inference for state space models using sequential Monte Carlo methods, but will also point to the use of SMC for ABC inference in other settings. I will also give a brief overview of some of the more recent work in the ABC literature, concluding with a look at performing ABC inference for spatial point process models, with applications in rainforest ecology.

Yanbo Tang – Practical Use of Synthetic Likelihood

 We present the synthetic likelihood approach to modeling complex models with intractable likelihoods. This was first proposed by Wood (2010) for ecological time series, but this method can be generalized to any model which is cheap to simulate from. The key idea is to replace the initial likelihood by an approximate likelihood constructed from well-chosen summary statistics, which are assumed to be asymptotically normal. We discuss the approach from a practical and theoretical point of view and consider the some of its benefits and flaws from an end-user perspective. We conclude by briefly introducing the Bayesian Synthetic Likelihood, which is the Bayesian extension to the initial approach.

Lucas Makinen – Information-maximising neural networks

How much information is embedded in your dataset, and can it be extracted ? Many sciences have become big-data-driven–cosmological surveys capture rich images to describe cosmic structure using millions of pixels, and online social networks can have billions of high-dimensional nodes–neither of which can be interrogated without massive compression. This poses a problem for simulation-based inference: how can we best compare simulations to reality ?

Information Maximising Neural Networks offer a way to compress massive datasets down to (asymptotically) lossless summaries that contain the same information as a full sky survey or network graph, as well as quantify the information content of an unknown distribution. We will show how to optimize these networks and use them for cheap simulation-based inference for a variety of data structures.

Session 2:

Mikael Mieskolainen  – Simulation and AI-driven inference in high energy physics

 High energy physics measurements are heavily aided by highly sophisticated Monte Carlo simulations of elementary collision processes and various (sub-)detector responses. In this context, I will introduce different measurement philosophies and explain how modern AI-driven high-dimensional statistical inference or inverse problem algorithms can be used to address them. These topics include a range of areas, from differential scattering cross-section and particle decay rate measurements to model parameter inference. I will demonstrate how the computational complexity, diversity, and high dimensionality of physical observables pose specific challenges on the road towards new physics discoveries.

Austin Mroz – The role of computation in the discovery of synthesisable molecular materials

 Novel functional materials are urgently needed to help combat the major global challenges facing humanity, such as climate change and resource scarcity. Traditional materials discovery initiatives are founded on intuition-guided, “trial-and-error” processes, where small, iterative changes to chemical structure and experimental conditions are made by the researcher. This is significantly resource and time intensive; after each small modification, the molecule must be synthesized (often a trial-and-error process in itself) and properties measured. As a result, these workflows are associated with long timescales (~20 years) and high costs. Recently, computation has accelerated this process via both atomistic simulations and data-driven approaches. The Jelfs Group takes advantage of these computational tools to investigate and predict molecular material assembly and bulk properties, as well as generate novel structures using artificial intelligence for experimental realisation.

 Andreas Joergensen  – Efficient Bayesian inference for stochastic agent-based models 

Mathematical models can help researchers understand tumour growth and develop patient-specific treatments. For instance, one might draw on agent-based models that track individual cells or cell clusters to study cancer on a microscopic level. Agent-based models successfully recover the observed macroscopic growth patterns of tumours and provide insights into the underlying mechanisms. However, agent-based models are computationally expensive, which might render parameter inferences through Bayesian sampling schemes insurmountable. Moreover, the models are stochastic, i.e. the model predictions change if the simulations are repeated. Indeed, this complication arises for many biological systems and can become yet another stumbling block for inference algorithms. In the talk, I will discuss how machine learning can help us overcome computational constraints while still accounting for the intrinsic stochasticity of biological systems.

Clotilde Cucinotta –Towards a Realistic Modelling of Electrified Interfaces the Nanoscale

In this talk In this presentation, I will discuss the challenges associated with simulating electrified interfaces at the nanoscale from first principles. Specifically, I will focus on simulating the effects of an applied potential to an electrochemical cell, using realistic models for the charged electrode-electrolyte interface. I will also share recent progress made in simulating the double layer of the fundamental Pt-water interface and its response to changes in the potential applied to the cell [1]. To achieve this, we have developed a general ab initio electrode-charging approach.

[1] R. Khatib, A. Kumar, S. Sanvito, M. Sulpizi and C. S. Cucinotta*, The nanoscale structure of the Pt-water double layer under applied potential revealed, 2021, 391, 138875