AI, Uncertainty & Simulation

About the Workshop

Bayesian Methods and
Neural Inference

Modern AI systems increasingly interact with scientific simulators, high-dimensional data, and safety-critical decisions. Yet many AI models remain poorly calibrated, overconfident, and unable to reason coherently under uncertainty.

This graduate workshop develops the mathematical and computational foundations of AI under uncertainty, beginning with Bayesian inference and building toward modern simulation-based inference (SBI) and neural emulation.

Participants will learn how probabilistic reasoning, approximate inference, and surrogate modelling form the backbone of scalable and reliable AI systems — especially when explicit likelihoods are unavailable but simulators are abundant.

The week culminates in hands-on JEDI-style team research projects applying the techniques covered in the lectures in a variety of fields such as astronomy, medicine and engineering.

"Uncertainty as infrastructure for reliable AI."

Core Themes

AI Beyond Point Prediction

Move from deterministic outputs to full posterior distributions

Uncertainty as Infrastructure

Principled probabilistic reasoning at scale

Inference as Inversion

Learning from generative processes and simulators

Neural Methods, Statistical Principles

Deep learning grounded in Bayesian foundations

Why This Workshop?

Many AI methods can be understood as approximate Bayesian inference at scale. Simulation-based inference generalises this paradigm to settings where likelihoods are intractable but forward simulators exist.

This workshop provides a unified conceptual framework connecting:
Bayesian statistics → probabilistic deep learning → likelihood-free inference → neural emulation → scalable scientific AI.

Learning Outcomes

What You Will Learn

Participants will develop a coherent understanding of the foundations to blend rigorous statistics with AI, with an emphasis on conceptual clarity, mathematical grounding, and computational practice.

Bayesian inference as a framework for scientific learning
Posterior geometry and hierarchical modelling
Variational inference and amortized neural inference
Generative models as probabilistic programs
Likelihood-free inference and simulation-based inference (SBI)
Neural density estimation and posterior estimation
Gaussian process and neural emulators
Calibration, proper scoring rules, and reliability diagnostics
Active learning and adaptive simulation strategies

Programme

Four Days of Deep Engagement

Bayesian Foundations for AI

Probability as inference, posterior reasoning, hierarchical models, predictive checks, decision theory, and MCMC intuition.

09:00 – 10:30

Bayesian Inference from First Principles

Probability as epistemic uncertainty. Bayes' theorem. Prior elicitation and posterior reasoning. Conjugate families and the Beta-Binomial model.

11:00 – 12:30

Hierarchical Models & Posterior Geometry

Hierarchical Bayesian models. Shrinkage and partial pooling. Understanding posterior landscapes. Predictive distributions and model checking.

14:00 – 15:30

MCMC: Metropolis, Gibbs, and HMC

Markov Chain Monte Carlo fundamentals. Metropolis-Hastings, Gibbs sampling, Hamiltonian Monte Carlo. Diagnostics: R-hat, effective sample size, trace plots.

16:00 – 17:30

Hands-On Lab: Bayesian Workflow in Stan

Practical session: implementing a complete Bayesian workflow with Stan. Prior/posterior predictive checks. Model comparison.

Neural Inference & Probabilistic Deep Learning

Variational inference, latent variable models, neural density estimation, amortized inference, uncertainty in deep learning.

09:00 – 10:30

Variational Inference & the ELBO

Evidence lower bound (ELBO). Mean-field approximations. Reparameterization trick. Automatic differentiation variational inference (ADVI).

11:00 – 12:30

Latent Variable Models & VAEs

Variational autoencoders (VAEs). Amortized inference with encoder-decoder architectures. Normalizing flows for flexible posteriors.

14:00 – 15:30

Neural Density Estimation

Conditional density networks. Mixture density networks. Masked autoregressive flows (MAF). Neural spline flows. Coupling layers and invertible transformations.

16:00 – 17:30

Uncertainty in Deep Learning

Bayesian neural networks. Deep ensembles. MC Dropout. Laplace approximation at scale. Epistemic vs aleatoric uncertainty.

Simulation-Based Inference & Emulation

Inference without likelihoods, ABC, neural posterior estimation, emulator construction, surrogate modelling, active simulation design.

09:00 – 10:30

Likelihood-Free Inference: From ABC to SBI

Approximate Bayesian Computation (ABC). Summary statistics and distance functions. Neural posterior estimation (NPE), likelihood estimation (NLE), and ratio estimation (NRE).

11:00 – 12:30

Neural SBI in Practice

The SBI toolkit. Sequential neural posterior estimation. Embedding networks for high-dimensional observations. Applications to cosmology and neuroscience.

14:00 – 15:30

Gaussian Processes & Neural Emulators

GP regression and kernel design. Uncertainty-aware surrogate models. Neural network emulators. Deep kernel learning. Multi-fidelity emulation.

16:00 – 17:30

Calibration & Active Learning

Proper scoring rules. Expected calibration error (ECE). Reliability diagrams. Active learning for simulation budgets. Bayesian optimization and acquisition functions.

Advanced Research Session — Frontiers

Cutting-edge methods in neural SBI, high-dimensional emulation, adaptive simulation budgets, calibration, and scientific AI applications.

09:00 – 10:00

Keynote: Neural SBI at Scale

Invited keynote on amortized inference, scalable posteriors, and open challenges in simulation-based inference.

10:30 – 12:30

Research Talks & Methods

Three contributed talks: high-dimensional emulation for climate models, neural ABC for genomics, sequential experimental design with neural surrogates.

14:00 – 16:00

Collaborative Problem-Solving

Breakout groups tackle domain-specific challenges: climate model calibration, gravitational wave inference, drug discovery optimization.

16:00 – 17:00

Presentations & Closing

Groups report findings. Open discussion on future research directions. Closing remarks and networking.

Lecturers

Our lecturers and organising committee for AI, Uncertainty & Simulation 2026.

B.B

Lecturer.

Prof. Bruce Bassett

Wits University and Stellenbosch University

Artificial Intelligence, Bayesian Inference, Medical AI, Natural Language Processing

A.H

Lecturer.

Prof. Alan Heavens

Imperial College London

Cosmology, Bayesian Inference, Weak Lensing

A.J

Lecturer.

Prof. Andrew Jaffe

Imperial College London

Bayesian Inference, Cosmology, Cosmic Microwave Background

N.O

Lecturer.

Dr. Nadeem Oozeer

SARAO

Data Science, Machine Learning, Radio Astronomy, Spectrum Pollution

A.R

Lecturer.

Dr. Amy Rouillard

Wits Health Consortium and Wits MIND Institute

AI for Healthcare, Quantum Algorithm Design, Data Science

Organising Committee

Linda Camara (Wits University)
Daniesha Govender (Wits University)
Robert Lees (Wits University)

Audience

Who Should Attend

This workshop is designed for:

PhD students and postdoctoral researchers
MSc students in statistics, ML, applied mathematics, physics, computational biology, or engineering
Researchers working with complex simulators (climate, cosmology, biology, epidemiology)
Anyone interested in building reliable AI systems under uncertainty

Prerequisites: Familiarity with statistics, linear algebra and coding is expected. Coding examples will be in Python. Prior exposure to Bayesian inference or deep learning is helpful but not required.

You Will Leave With

Mathematical insight into modern probabilistic AI

Practical tools: Stan, PyMC, SBI toolkits, GPyTorch

A unified framework connecting Bayesian stats → neural SBI → emulation

Build a global network of researchers interested in the frontiers of AI in science.

Applications

Apply to Attend

What to Submit

Applications should include:

A short statement of motivation (maximum 500 words)
A current CV or academic résumé
A brief description of your research background and how this workshop relates to your interests
A letter of agreement from your supervisor
A letter of reference sent directly to us by the application deadline (please indicate the referee's name in your application)

Financial Support

Full scholarships covering travel and accommodation are available for African participants. As participation in the workshop is limited, applications will be reviewed and selected on a competitive basis.

Apply via Online Form

Key Dates

Application Deadline31 May 2026

Decisions Communicated30 June 2026

Workshop Dates16–20 November 2026

FAQ

Frequently Asked Questions

No. We start from first principles. However, some familiarity with probability theory, linear algebra, and Python programming is expected.

Yes. Each day includes practical labs with Python. We'll use Stan, PyMC, the SBI toolkit, GPyTorch, and other modern probabilistic programming tools.

Full scholarships are available for African students.

A laptop with Python 3.10+ and conda/pip installed. We'll provide setup instructions and environment files in advance.

This may be possible for lectures only; however, remote attendance is not available for group work.

AI, Uncertainty& Simulation

Bayesian Methods andNeural Inference

Core Themes

AI Beyond Point Prediction

Uncertainty as Infrastructure

Inference as Inversion

Neural Methods, Statistical Principles

Why This Workshop?

What You Will Learn

Four Days of Deep Engagement

Lecturers

Who Should Attend

You Will Leave With

Apply to Attend

What to Submit

Financial Support

Key Dates

Frequently Asked Questions

AI, Uncertainty
& Simulation

Bayesian Methods and
Neural Inference