ML Nosh

ML Nosh Seminar Series

This is a community-building initiative aimed at connecting and strengthening Oxford’s diverse and multi-departmental research community in Artificial Intelligence and Machine Learning.

Each session features a concise, high-level 15-minute presentation, followed by an informal lunch to to encourage open discussions in a relaxed atmosphere.

The seminar series takes place bi-weekly. To sign up for the mailing list, please fill in this form.

Organisers: Andrei Constantin, Ard Louis and Shivaji Sondhi

Past events

Monday, 10 March 2025, 13:00: Atılım Güneş Baydin (Department of Computer Science)

Venue: Martin Wood Lecture Theatre, Department of Physics

Probabilistic and Differentiable Programming: A New Paradigm for Scientific Simulation
This talk explores how machine learning is transforming scientific simulation through probabilistic and differentiable programming. By treating complex simulators as probabilistic programs, we enable automated Bayesian inversion for discovery in high-energy physics and astrophysics. We also discuss the use of surrogate models and automatic differentiation to accelerate simulation-based inference for applications such as heliophysics analyses and molecular complexity prediction. These approaches, leveraging distributed computing and amortized inference, significantly enhance efficiency and scalability while fostering interdisciplinary collaboration. 

Monday, 24 February 2025, 13:00: Dominik Lukeš (Lead Business Technologist at the AI/ML Support Competency Centre)

Venue: Martin Wood Lecture Theatre, Department of Physics

Metaphors of Reason and the Large Language Models: Revisiting the Moravec Paradox in the Face of a Contentious Issue 
This talk examines how our inherited metaphorical frameworks for reasoning illuminate contemporary debates about the capabilities of Large Language Models. Through a critical analysis of dominant metaphors - from reasoning as symbolic manipulation or knowledge retrieval to reasoning as intuitive pattern recognition - the presentation will show how taking the Moravec Paradox seriously can offer fresh analytical perspectives on the apparent contradictions in contrasting machine and human intelligence capabilities. Examining these metaphors will also help us reframe some persistent puzzles regarding human learning of abstract concepts. 

Monday, 10 February 2025, 13:00: Francesco Mori (Department of Physics)

Venue: Martin Wood Lecture Theatre, Department of Physics

Optimal learning strategies via statistical physics and control theory 
Training machine learning models involves optimising strategies such as adaptive hyper-parameters, data selection, and dynamic architectures to improve performance. However, these approaches often rely on heuristic trial-and-error methods, which lack a solid theoretical foundation and may lead to suboptimal results. In this talk, I will present an integrated framework that combines dimensionality-reduction techniques from statistical physics with control-theoretic methods to derive optimal training strategies for prototypical learning problems. Specifically, I will focus on shallow neural networks trained with online stochastic gradient descent. By deriving closed-form ordinary differential equations that describe the evolution of low-dimensional order parameters, I will provide a complete characterisation of the network’s performance during training. Applying control theory to this reduced description, it is then possible to identify training strategies, including learning rate schedules and curricula, that enhance generalisation. 

Monday, 27 January 2025: Sam Staton (Department of Computer Science)

Venue: Martin Wood Lecture Theatre, Department of Physics

Probabilistic Programming and ML
Probabilistic programming is a method for writing statistical models by writing programs. I'll give an introduction to probabilistic programming, and link to recent ideas in AI, such as the ARIA Safeguarded AI programme, as well as to my own work on programming language foundations. 

Monday, 2 December 2024: Andre Lukas (Department of Physics)

Venue: Lindemann Lecture Theatre, Department of Physics 

Machine Learning and String Theory
I will introduce some properties of string theory and explain why certain problems within this theory lend themselves to machine learning methods. This includes exploring the large model spaces of string theory which, in practice, requires solving certain Diophantine equations, and solving partial differential equations, typically on complicated spaces, such as Calabi-Yau manifolds. I will present some explicit examples and explain how methods such as reinforcement learning, genetic algorithms and neural network techniques to solve differential equations can be applied to these problems.

Monday, 18 November 2024: Michael Osborne (Department of Engineering)

Venue: Lindemann Lecture Theatre, Department of Physics

Insuring Emerging Risks From AI
We examine the implications of recent progress in artificial intelligence (AI) for liability regimes and insurance markets within the United States. We argue that the insurance industry faces both a potential decline in traditional markets like auto insurance and emerging growth opportunities in AI agent and cybersecurity coverage. The report advocates for targeted reforms in liability laws, proposing a nuanced approach that may ease regulations for demonstrably-safer technologies, such as future autonomous vehicles, whilst strengthening oversight for AI agents and cyber risks. These proposed changes would significantly impact the insurance sector, necessitating the development of new actuarial methodologies to quantify complex AI-related risks and to potentially underwrite a broader range of liabilities. We conclude that the insurance industry has a pivotal role to play in managing AI-related risks, fostering responsible innovation, and ensuring that the benefits of AI are broadly shared across society.