Events

Date: 17 September 2026   Time: 09:45 - 16:50

Location: Fogg Lecture Theatre, Fogg Building, Mile End Campus

Following a successful first symposium in April, which drew more than 40 attendees and sparked several new collaborations, we're pleased to announce the second "AI: Brains and Bits" symposium. The series brings together researchers from Neuroscience, Electronic Engineering, Computer Science, Mathematics, Physics and beyond.
This session's programme features short talks from:

• Dilay Erceli (University of Oxford, Department of Pharmacology)
• Gabriel Bontemps
• Diana Benavides Prado (EECS)
• Paul Curzon (EECS)
• Emma Stewart (SBBS)
• David Mguni (EECS)
• Guifen Chen (SBBS)

• Ian McFadden (SBBS)

• Enrico Camporeale (SPCS)

Each speaker will introduce their work with an eye to finding new collaborators in the room, so expect open questions and pointers to the kind of expertise they're looking for, alongside their results. There will also be a series of one-minute pitches and small group discussion.

Spaces are limited, so please sign up by 8 September using the link below so we can arrange catering.
[SIGN-UP HERE]

Coffee and pastries from 09:45, welcome at 10:15. Please arrive by 10:00 if you can.

We look forward to seeing you there!

Abstracts

Learning Whom to Trust: Decision-Generated Credibility in Social Learning

Gabriel Bontemps, Queen Mary University of London

Social interaction can improve collective learning but also amplify early mistakes when credibility is generated by the sender's own decision process. We model agents whose choices, decision times, and confidence arise jointly through a drift-diffusion process, with confidence determining the weight of social influence and learning. The resulting social dynamics yield an analytical amplification threshold and show how cross-community permeability regulates disagreement. The model identifies when endogenous credibility supports collective correction and when it instead reinforces social error.

Spatial maps in "hard mode"

Guifen Chen, presenting online

How does an adult brain build a new internal representation of the world? Using spatial learning as a model, I show how hippocampal place representations emerge rapidly in a new environment, while entorhinal grid representations develop more gradually. My aim is to extend this work to understand how different sensory inputs combine to shape this process. I am particularly interested in collaborating with researchers in AI and computational modelling to explore the common principles of how new knowledge is incorporated into established neural circuits and how artificial agents build stable internal representations from experience.

Disentangling Latent Computations in Mouse Reversal Learning: Linking Behaviour and Neural Dynamics

Dilay Ercelik, University of Oxford, Department of Pharmacology

In reversal learning, animals must use recent cues, choices, and outcomes to infer when task rules have changed. I will present a Disentangled RNN (Miller et al., 2024) framework for modelling trial-level mouse lick/no-lick behaviour choices and calcium-derived neural features (i.e., behavioural and neural targets). The aim is to learn compact candidate latent variables that predict behaviour while remaining interpretable in relation to inferred rule, choice persistence, error sensitivity, uncertainty, and adaptation after reversals. I will discuss how comparing behaviour-only and joint behaviour-neural models can help identify which latent computations are stable across mice or suggestive of individual differences.

Inference in human perception and eye movements

Emma Stewart, School of Biological and Behavioural Sciences

Human visual experience is constrained by three key factors: 1) the information that hits the retina is an imperfect two-dimensional retinal projection of the world, which is 2) sampled non-uniformly via inherently biased eye movements, and 3) only a small portion of the world at a time is viewed in high-resolution central vision. Despite these limitations, we perceive a seamless percept of the world around us. I will discuss some of the mechanisms that humans use to overcome these limitations, by using inference, statistical learning, and prediction. While human perception is often modelled using machine learning, for the most part these models overlook the constraints inherent in vision. I am interested in collaborating with researchers in ML and robotics to see how accounting for human visual limitations affects predictions in humans and AI.

Beyond Static Models: Foundations and Open Frontiers of Continual Learning

Diana Benavides Prado

Artificial intelligence has achieved extraordinary success through static training regimes, yet true autonomy requires continuous learning and adaptation to an ever-evolving world. Continual machine learning seeks to bridge this gap by enabling agents to sequentially acquire, refine, and retain knowledge throughout their operational lifespan. This talk highlights key facets of continuous adaptation, including catastrophic forgetting and knowledge transfer, alongside some of our work leveraging transfer dynamics to achieve both retention and backward improvement. We examine theoretical foundations, such as the fundamental stability-plasticity dilemma in balancing memory preservation with model adaptability, as well as computational contrasts with human cognitive mechanisms like memory consolidation and targeted unlearning. We also reflect on the profound philosophical questions raised by self-improving agents that autonomously refine their own internal representations over time. Examining these foundational principles and open frontiers across science and engineering will help shape the next generation of lifelong learning AI.

The challenge of rare and extreme events: the space weather use-case

Enrico Camporeale

Rare and extreme events pose a fundamental challenge for machine learning: the events we care about most are often those for which we have the least data. Using space weather as a case study, I will discuss this challenge and introduce PARIS (Pruning Algorithm via the Representer theorem for Imbalanced Scenarios), a framework for improving the prediction of rare targets by strategically selecting training data.

Learning the Geometry of Intelligence

David Mguni

Modern AI systems learn to perform tasks, but often provide little insight into the underlying structure that makes those tasks difficult or determines which reasoning strategies succeed. In this talk I will present a research programme that aims to learn the hidden geometry of intelligent behaviour itself. Using examples from game theory, optimisation and large language models, I will show how representation learning, meta-learning and active experimentation can reveal latent structure that supports prediction, adaptation and failure discovery. I will conclude by discussing how these ideas point towards AI systems that build increasingly scientific models of their own behaviour.

Building a cross-centre research group to map layers of perception in ecology

Ian McFadden

Machine perception across vision, audio and other senses is rapidly advancing, as are the many threats to global ecosystems, species and the ecological interactions that weave them together. Now is the time to build ecological general intelligence systems that can understand, navigate and restore the many layers of complexity inherent in nature. I will introduce the Eco-AI group, consisting of lecturers and PhD students working across the Centre for Biodiversity and Sustainability and the Centre for Multimodal AI, whose goal is to map ecological complexity conceptually, visually and spatially. We hope to grow the collaborative network of the group through mutually beneficial research.

What legitimates AI use?

Paul Curzon

The sociology framework Legitimation Code Theory (LCT) provides ways to investigate what legitimates knowledge practice including how people are afforded legitimation in social arenas. With the widespread use of AI and agentic AI such questions can be asked of AI actors. In past research, interviewing patients with chronic illness about use of AI to make treatment decisions, a clear message was that it would only be a good idea with a human in the loop. AIs were not seen as legitimate replacements for human clinicians. I am interested in exploring what does give AI legitimation. I am also developing AI systems to do LCT analysis and exploring the use of LCT to underpin explainable AI.