Eleanor Spens

I’m a neuroscientist working on how our brains learn predictive models of the world offline, how frontal cortex develops useful abstractions, and how this relates to recent advances in AI. Before joining St John’s, I was a postdoc in Tim Behrens’ lab in Oxford, following a PhD with Neil Burgess at UCL (modelling systems consolidation as the training of generative models in cortex on replayed memories). Before that, I worked in machine learning in the UK Civil Service for a few years. I studied Physics and Philosophy at Oxford for my first degree.

The brain develops a hierarchy of abstractions by learning to compress experience. This limits the amount of information that must be stored in episodic memory (to remember past events) or in working memory (to plan future behaviour).

My previous work has mostly explored how we learn from episodic memories: as hippocampal memories are replayed over the course of systems consolidation, they update a generative model of the world that supports memory as well as reasoning about future situations. Even right after an experience, remembering involves imagining the past based on concepts, combining some stored details with expectations about what happened.

Our knowledge of the world has complex hierarchical structure. When watching a TV show, one keeps track of information over varying timescales, e.g. the gist of the series so far, the plot of the current episode, and the sensory stimuli in the last few seconds. Such conceptual hierarchies allow us to understand and plan over long timescales without representing every detail at once. That is, representing long and complex events efficiently by compressing them as much as possible allows storage despite limits on episodic memory capacity and reasoning despite limits on working memory capacity.

My current research involves building computational models in which events are represented at multiple levels of abstraction, considering a range of stimuli from trajectories in mazes to narratives, and comparing their predictions to neural data. The central idea is that the brain should represent information at the level of abstraction most useful for the current problem: items should be represented in detail in working memory when they deviate from predictions, but compressed when predictable.

Recent publications

https://www.nature.com/articles/s41562-023-01799-z: Spens, E., & Burgess, N. (2024). A generative model of memory construction and consolidation. Nature Human Behaviour, 8(3), 526–543.

https://www.nature.com/articles/s41467-026-74357-6: Spens, E., & Burgess, N. (2026). Hippocampo-neocortical interaction as compressive retrieval-augmented generation. Nature Communications, 17, 7971.

https://openreview.net/pdf?id=BFW1fkB8ck: Spens, E., Burgess, N., & Behrens, T. (2025). Modelling the control of offline processing with reinforcement learning. Advances in Neural Information Processing Systems, 38, 163024–163046.