Schedule 2026/27

Here is the schedule for the 2026/27 academic year.

When: 15:00–16:00 UK local time, Thursdays. Venue: Charles Carter, A02.

Michaelmas Term

Date Speaker Title Links
29 Oct 2026 Mher Safaryan Lancaster University On Gradient Estimation for Quantization-Aware Training

Quantization-aware training (QAT) enables low-bit inference with reduced memory and computational costs. However, achieving competitive accuracy compared to full-precision training is challenging, particularly at very low bit widths. A central difficulty lies in gradient mismatch: quantization introduces non-differentiable operations, typically handled using the straight-through estimator (STE), which does not fully account for the effects of quantization on optimisation. In this talk, I will revisit gradient estimation in QAT from an optimisation perspective. I will introduce a multi-objective formulation that balances loss minimisation with adherence to quantization constraints and discuss how this viewpoint motivates alternative gradient estimators that incorporate quantization error into the training dynamics. I will discuss its theoretical foundations, including Pareto optimality and a convergence result, and its efficient implementation with standard optimisers alongside experimental results on low-bit fine-tuning and pre-training of large language models.
5 Nov 2026 Alexandra Borkowski King's College London Acceleration in Underdamped Langevin Dynamics via Fenchel Games

In this talk, we formulate sampling as a two-player game and identify Underdamped Langevin Dynamics as coupled strategies of the players. This perspective builds on similar constructions for accelerated gradient flows in ℝᵈ. A cost function involving fictitious competitors enables us to control the KL divergence between the positional marginal and the target distribution π, which is assumed to be log-concave or strongly log-concave. In this way, we obtain accelerated convergence rates compared to those of the corresponding Overdamped Langevin Dynamics.
12 Nov 2026 Matthias Sachs Lancaster University
19 Nov 2026 Gavin Kerrigan University of Oxford
26 Nov 2026 Gabriel Stoltz CERMICS, École des Ponts & Inria Paris Nonequilibrium stochastic dynamics: Error estimates and variance reduction

Equilibrium properties in statistical physics are obtained by computing averages with respect to Boltzmann-Gibbs measures, sampled in practice using ergodic dynamics such as the Langevin dynamics. Some quantities however cannot be computed by simply sampling the Boltzmann-Gibbs measure, in particular transport coefficients, which relate the current of some physical quantity of interest with the forcing needed to induce it. For instance, a temperature difference induces an energy current, the proportionality factor between these two quantities being the thermal conductivity. From an abstract point of view, transport coefficients can also be considered as some form of sensitivity analysis with respect to an added forcing to the baseline dynamics.

There are various numerical techniques to estimate transport coefficients, which all suffer from large errors, in particular large statistical errors. I will review the most popular methods, namely the Green–Kubo approach where the transport coefficient is rewritten as some time-integrated correlation function, and the approach based on longtime averages of the stochastic dynamics perturbed by an external driving (so-called nonequilibrium molecular dynamics). I will make precise in each case the various sources of errors, in particular the bias related to the time discretization of the underlying continuous dynamics, and the variance of the associated Monte Carlo estimators. I will also briefly present some recent alternative techniques to estimate transport coefficients.
3 Dec 2026 Paula Cordero Encinar Imperial College London
Date Speaker Title Links
29 Oct 2026 Mher Safaryan Lancaster University On Gradient Estimation for Quantization-Aware Training

Quantization-aware training (QAT) enables low-bit inference with reduced memory and computational costs. However, achieving competitive accuracy compared to full-precision training is challenging, particularly at very low bit widths. A central difficulty lies in gradient mismatch: quantization introduces non-differentiable operations, typically handled using the straight-through estimator (STE), which does not fully account for the effects of quantization on optimisation. In this talk, I will revisit gradient estimation in QAT from an optimisation perspective. I will introduce a multi-objective formulation that balances loss minimisation with adherence to quantization constraints and discuss how this viewpoint motivates alternative gradient estimators that incorporate quantization error into the training dynamics. I will discuss its theoretical foundations, including Pareto optimality and a convergence result, and its efficient implementation with standard optimisers alongside experimental results on low-bit fine-tuning and pre-training of large language models.
5 Nov 2026 Alexandra Borkowski King's College London Acceleration in Underdamped Langevin Dynamics via Fenchel Games

In this talk, we formulate sampling as a two-player game and identify Underdamped Langevin Dynamics as coupled strategies of the players. This perspective builds on similar constructions for accelerated gradient flows in ℝᵈ. A cost function involving fictitious competitors enables us to control the KL divergence between the positional marginal and the target distribution π, which is assumed to be log-concave or strongly log-concave. In this way, we obtain accelerated convergence rates compared to those of the corresponding Overdamped Langevin Dynamics.
12 Nov 2026 Matthias Sachs Lancaster University
19 Nov 2026 Gavin Kerrigan University of Oxford
26 Nov 2026 Gabriel Stoltz CERMICS, École des Ponts & Inria Paris Nonequilibrium stochastic dynamics: Error estimates and variance reduction

Equilibrium properties in statistical physics are obtained by computing averages with respect to Boltzmann-Gibbs measures, sampled in practice using ergodic dynamics such as the Langevin dynamics. Some quantities however cannot be computed by simply sampling the Boltzmann-Gibbs measure, in particular transport coefficients, which relate the current of some physical quantity of interest with the forcing needed to induce it. For instance, a temperature difference induces an energy current, the proportionality factor between these two quantities being the thermal conductivity. From an abstract point of view, transport coefficients can also be considered as some form of sensitivity analysis with respect to an added forcing to the baseline dynamics.

There are various numerical techniques to estimate transport coefficients, which all suffer from large errors, in particular large statistical errors. I will review the most popular methods, namely the Green–Kubo approach where the transport coefficient is rewritten as some time-integrated correlation function, and the approach based on longtime averages of the stochastic dynamics perturbed by an external driving (so-called nonequilibrium molecular dynamics). I will make precise in each case the various sources of errors, in particular the bias related to the time discretization of the underlying continuous dynamics, and the variance of the associated Monte Carlo estimators. I will also briefly present some recent alternative techniques to estimate transport coefficients.
3 Dec 2026 Paula Cordero Encinar Imperial College London

Michaelmas Term

Date Speaker Title Links
22 Oct 2026 Shreya Sinha Roy Lancaster University Prequential posteriors: theory and application

Data assimilation is a fundamental task in updating forecasting models upon observing new data, with applications ranging from weather prediction to online reinforcement learning. Deep generative forecasting models (DGFMs) have shown excellent performance in these areas, but assimilating data into such models is challenging due to their intractable likelihood functions. This limitation restricts the use of standard Bayesian data assimilation methodologies for DGFMs. To overcome this, we introduce prequential posteriors, based upon a predictive-sequential (prequential) loss function; an approach naturally suited for temporally dependent data which is the focus of forecasting tasks. Since the true data-generating process often lies outside the assumed model class, we adopt an alternative notion of consistency and prove that, under mild conditions, both the prequential loss minimizer and the prequential posterior concentrate around parameters with optimal predictive performance. For scalable inference, we employ easily parallelizable wastefree sequential Monte Carlo (SMC) samplers with preconditioned gradient-based kernels, enabling efficient exploration of high-dimensional parameter spaces such as those in DGFMs. We validate our method on both a synthetic multi-dimensional time series and a real-world meteorological dataset; highlighting its practical utility for data assimilation for complex dynamical systems. We also apply the proposed method for policy learning in the context of model-based reinforcement learning and achieve rapid identification of high-reward policies in complex tasks.
15 Oct 2026 Jixiang Qing Lancaster University Multi-stage Bayesian optimisation under unknown stopping constraints

Many real-world black-box processes involve sequential decision-making under stage-wise constraints, where violating an unknown constraint can terminate the entire run. Learning to make optimal decisions in such settings with limited evaluations is therefore an important challenge. In this talk, I will present our multi-stage Bayesian optimisation framework for tackling this problem. I will show how the problem can be formulated as a Markov decision process and approached through dynamic programming, leading to theoretical guarantees on both cumulative failures and regret. Finally, I will demonstrate the empirical performance of our approach on synthetic benchmarks and application-driven optimisation problems.
Date Speaker Title Links
22 Oct 2026 Shreya Sinha Roy Lancaster University Prequential posteriors: theory and application

Data assimilation is a fundamental task in updating forecasting models upon observing new data, with applications ranging from weather prediction to online reinforcement learning. Deep generative forecasting models (DGFMs) have shown excellent performance in these areas, but assimilating data into such models is challenging due to their intractable likelihood functions. This limitation restricts the use of standard Bayesian data assimilation methodologies for DGFMs. To overcome this, we introduce prequential posteriors, based upon a predictive-sequential (prequential) loss function; an approach naturally suited for temporally dependent data which is the focus of forecasting tasks. Since the true data-generating process often lies outside the assumed model class, we adopt an alternative notion of consistency and prove that, under mild conditions, both the prequential loss minimizer and the prequential posterior concentrate around parameters with optimal predictive performance. For scalable inference, we employ easily parallelizable wastefree sequential Monte Carlo (SMC) samplers with preconditioned gradient-based kernels, enabling efficient exploration of high-dimensional parameter spaces such as those in DGFMs. We validate our method on both a synthetic multi-dimensional time series and a real-world meteorological dataset; highlighting its practical utility for data assimilation for complex dynamical systems. We also apply the proposed method for policy learning in the context of model-based reinforcement learning and achieve rapid identification of high-reward policies in complex tasks.
15 Oct 2026 Jixiang Qing Lancaster University Multi-stage Bayesian optimisation under unknown stopping constraints

Many real-world black-box processes involve sequential decision-making under stage-wise constraints, where violating an unknown constraint can terminate the entire run. Learning to make optimal decisions in such settings with limited evaluations is therefore an important challenge. In this talk, I will present our multi-stage Bayesian optimisation framework for tackling this problem. I will show how the problem can be formulated as a Markov decision process and approached through dynamic programming, leading to theoretical guarantees on both cumulative failures and regret. Finally, I will demonstrate the empirical performance of our approach on synthetic benchmarks and application-driven optimisation problems.