Tue 10 Jun 11:00: Context, Computation, and Continuity: Neural Mechanisms of Memory and Decision-Making
How does the brain achieve both stability and flexibility in behavior? In this journal club, we will explore two recent studies that illuminate distinct yet interconnected neural mechanisms underlying long-term motor memory and flexible decision-making. The first paper (Kim et al., Nature, 2024) demonstrates that motor memories are encoded in a combinatorial, context-specific manner in the motor cortex of mice. Using long-term two-photon imaging, the authors show that new motor skills are acquired without overwriting old ones, as new preparatory activity patterns emerge in parallel across contexts—offering a robust mechanism for continual learning.
The second paper (Pagan et al., Nature, 2024) investigates how individual variability shapes context-dependent decision-making in rats. The authors develop a behavioral paradigm and theoretical framework revealing three distinct dynamical strategies for evidence accumulation, all capable of supporting flexible behavior. Strikingly, different individuals express different combinations of these strategies, despite similar performance, highlighting substantial neural and computational diversity.
Optionally, we will also discuss findings from a third study (Mishchanchuk et al., Science, 2024), which reveals how the ventral hippocampus encodes abstract contextual states critical for hidden state inference. This study complements the others by highlighting the importance of hippocampal representations in decision-making based on latent contexts. Together, these studies provide a compelling picture of how the brain balances flexibility and stability through context-specific encoding, diverse computational strategies, and abstract contextual inference—shedding light on the neural basis of learning, memory, and cognition.
- Speaker: Máté Lengyel; Guillaume Hennequin
- Tuesday 10 June 2025, 11:00-12:30
- Venue: CBL Seminar Room, Engineering Department, 4th floor Baker building.
- Series: Computational Neuroscience; organiser: .
Tue 10 Jun 11:00: Context, Computation, and Continuity: Neural Mechanisms of Memory and Decision-Making
How does the brain achieve both stability and flexibility in behavior? In this journal club, we will explore two recent studies that illuminate distinct yet interconnected neural mechanisms underlying long-term motor memory and flexible decision-making. The first paper (Kim et al., Nature, 2024) demonstrates that motor memories are encoded in a combinatorial, context-specific manner in the motor cortex of mice. Using long-term two-photon imaging, the authors show that new motor skills are acquired without overwriting old ones, as new preparatory activity patterns emerge in parallel across contexts—offering a robust mechanism for continual learning. The second paper (Pagan et al., Nature, 2024) investigates how individual variability shapes context-dependent decision-making in rats. The authors develop a behavioral paradigm and theoretical framework revealing three distinct dynamical strategies for evidence accumulation, all capable of supporting flexible behavior. Strikingly, different individuals express different combinations of these strategies, despite similar performance, highlighting substantial neural and computational diversity. Optionally, we will also discuss findings from a third study (Mishchanchuk et al., Science, 2024), which reveals how the ventral hippocampus encodes abstract contextual states critical for hidden state inference. This study complements the others by highlighting the importance of hippocampal representations in decision-making based on latent contexts. Together, these studies provide a compelling picture of how the brain balances flexibility and stability through context-specific encoding, diverse computational strategies, and abstract contextual inference—shedding light on the neural basis of learning, memory, and cognition.
- Speaker: Máté Lengyel; Guillaume Hennequin
- Tuesday 10 June 2025, 11:00-12:30
- Venue: CBL Seminar Room, Engineering Department, 4th floor Baker building.
- Series: Computational Neuroscience; organiser: .
Thu 12 Jun 16:00: “Characterizing and preventing host cell entry of emerging RNA viruses” Please note the change of venue to: Max Perutz Lecture Theatre, MRC LMB
This Cambridge Immunology Network Seminar will take place on Thursday 12 June 2025, starting at 4:00-5:00pm
Speaker: Professor Thomas Bowden, Wellcome Centre for Human Genetics, Oxford
Title: “Characterizing and preventing host cell entry of emerging RNA viruses”
Host: Yorgo Modis, CITIID , Cambridge
Location: Max Perutz Lecture Theatre, MRC LMB
Refreshments will be available following the seminar.
Please note the change of venue to: Max Perutz Lecture Theatre, MRC LMB
- Speaker: Professor Thomas Bowden, Wellcome Centre for Human Genetics, Oxford
- Thursday 12 June 2025, 16:00-17:00
- Venue: Max Perutz Lecture Theatre, MRC LMB, Francis Crick Avenue, Cambridge CB2 0QH .
- Series: Cambridge Immunology Network Seminar Series; organiser: Ruth Paton.
Thu 12 Jun 14:00: Probabilistically Robust Decision Making for Uncertain Dynamical Systems
Typically, how much do we know about an uncertain dynamical system (UDS) matters a lot when we want to control them. Aiming to accurately capture the evolution of such UDS is impossible as true system uncertainties cannot be captured exactly. Lack of exact system knowledge increases the difficulty in estimating the limits of the uncertain system’s performance. As a result, we often seek to control such UDS such that the resulting control decisions from Robust Decision Making (RDM) paradigms render the UDS insensitive to what we don’t know about them. However, nature can violate the assumptions that the RDM module assume for the system uncertainties with small probability. Controlling UDS under such unforeseen events necessitate the addition of probabilistic rigour on top of the existing RDM approaches. In this talk, I shall propose a Probabilistic RDM (PRDM) approach using the uncertain gap between the dynamical system models (with and without the uncertainty) induced by appropriate distance metric. The proposed framework will allow us to analyse the potential performance degradation of a control action on an UDS when such rare violation events occur. The fertile nature of the probabilistic robust control research area will be highlighted using a list of interesting future research directions.
The seminar will be held in JDB Seminar Room, Department of Engineering, and online (zoom): https://newnham.zoom.us/j/92544958528?pwd=YS9PcGRnbXBOcStBdStNb3E0SHN1UT09
- Speaker: Venkatraman Renganathan, Cranfield University
- Thursday 12 June 2025, 14:00-15:00
- Venue: JDB Seminar Room, Department of Engineering and online (Zoom).
- Series: CUED Control Group Seminars; organiser: Fulvio Forni.
Fri 13 Jun 15:00: Title to be confirmed
Abstract not available
- Speaker: Allan McRobie, University of Cambridge
- Friday 13 June 2025, 15:00-16:00
- Venue: CivEng Seminar Room (1-33) (Civil Engineering Building).
- Series: Engineering Department Structures Research Seminars; organiser: Shehara Perera.
Tue 10 Jun 14:00: Latent Concepts in Large Language Models
Large Language Models (LLMs) have achieved remarkable fluency and versatility—but understanding how they represent meaning internally remains a challenge. In this talk, we explore the emerging science of latent concepts in LLMs: the semantic abstractions implicitly encoded in their internal activations.
We examine how concepts—such as truthfulness, formality, or sentiment—can be represented as low-dimensional structures, discovered through training dynamics, and understood through the lens of linear algebra and associative memory. We discuss the implications for interpretability, robustness, and control, including how concepts can be steered at test time to adjust model behavior without retraining. Specifically, we explore empirical and theoretical evidence supporting the linear representation hypothesis, where such concepts correspond to vectors or affine subspaces, emerging naturally from training dynamics and next-token prediction objectives. We further show that LLMs behave as associative memory systems, retrieving outputs based on latent similarity rather than logical inference. This behavior underlies phenomena such as context hijacking, where semantically misleading prompts can bias the model’s response.
We introduce formal latent concept models that unify these ideas, describe conditions under which concepts are identifiable, and propose learning algorithms for extracting interpretable, controllable representations. We argue that such latent concept modeling offers a principled framework for bridging representation learning with interpretability and model alignment, and offers a promising path toward safer, more controllable, and more trustworthy AI.
- Speaker: Prof. Pradeep Ravikumar, Carnegie Mellon University
- Tuesday 10 June 2025, 14:00-15:00
- Venue: JDB Seminar Room, CUED.
- Series: Probabilistic Systems, Information, and Inference Group Seminars; organiser: Prof. Ramji Venkataramanan.
Wed 18 Jun 15:00: Title to be confirmed
Abstract not available
- Speaker: Max Ryabinin
- Wednesday 18 June 2025, 15:00-16:00
- Venue: Computer Lab, LT1.
- Series: Cambridge ML Systems Seminar Series; organiser: Sally Matthews.
Thu 12 Jun 11:30: Mixing and chemical transfers in particle clouds – implications following planetary impacts
At a late stage of its accretion, the Earth experienced high-energy planetary impacts. Following each collision, the metal core of the impactor sank as millimetric drops into a molten silicate magma ocean. The efficiency of chemical equilibration between these silicates and the metal core controlled the composition of the Earth controlled the initial temperature and composition of rocky planets, and hence the emergence of plate tectonics, the time when a solid inner core started to grow, or the driving of an early dynamo in the Earth’s core by exsolution of light elements.
In this talk I will present different experiments focusing on the interaction of settling particle clouds with their surrounding through entrainment, mixing and chemical reactions. I will first present experiments on inert clouds settling in a quiescent fluid. Then, I will discuss the implications of planetary rotation on the efficiency of chemical transfers inside particle clouds, largely disregarded despite the strong rotation rate of the proto-Earth that has been suggested by impact simulations.
- Speaker: Quentin Kriaa, Uni of Cambridge
- Thursday 12 June 2025, 11:30-12:30
- Venue: Open Plan Area, Institute for Energy and Environmental Flows, Madingley Rise CB3 0EZ.
- Series: Institute for Energy and Environmental Flows (IEEF); organiser: Catherine Pearson.
Thu 03 Jul 11:30: TBC
Abstract not available
- Speaker: Jason Furtney,
- Thursday 03 July 2025, 11:30-12:30
- Venue: Open Plan Area, Institute for Energy and Environmental Flows, Madingley Rise CB3 0EZ.
- Series: Institute for Energy and Environmental Flows (IEEF); organiser: Catherine Pearson.
Fri 13 Jun 14:00: Detecting Changes in Production Frontier and Beyond
In this talk, we first give a brief review of the nonparametric estimation problem of production frontier function, which concerns the maximum possible output given input levels and the efficiency of the firms. We then look at how multiple changes over time in the production frontier can be detected. By assuming that the frontier always shifts upwards over time, which is plausible thanks to the advance in technologies, we can detect changes in the frontier at the near-optimal rate under regularity conditions, irrelevant of the dimensionality of the input. This can be achieved by modifying and utilising the well-known Free Disposal Hull (FDH) or Data Envelopment Analysis (DEA) algorithm in different ways, depending on whether the shift is local or global. Finally, we discuss how the confidence intervals for both the location of the change and the frontier can be constructed. For the latter, we also illustrate how nuisance parameters in the limit distribution can be eliminated by taking advantage of the structure of the FDH , and demonstrate how this idea can be useful for other problems such as testing the shape of the frontier.
- Speaker: Yining Chen (LSE)
- Friday 13 June 2025, 14:00-15:00
- Venue: MR12, Centre for Mathematical Sciences.
- Series: Statistics; organiser: Qingyuan Zhao.
Thu 05 Jun 11:30: Unveiling complex transport processes in a large deep lake: From coastal upwelling to higher-mode internal waves
Water quality in lakes is closely linked to hydrodynamics and is often dominated by thermal stratification which limits the exchange between the upper layers (called the epilimnion) and the deeper layers (called the hypolimnion). Consequently, the vertical redistribution of biogeochemical tracers such as dissolved oxygen and nutrients by convective overturning during winter is a key process in annual lake cycles. In deep lakes, convective cooling often does not reach the deepest layers. Furthermore, convective cooling is weakening due to climate change, motivating a good understanding of (i) alternative deepwater renewal mechanisms, and (ii) deepwater dynamics in large deep lakes in general. Understanding deepwater dynamics is crucial because of the role deepwater currents play in mediating water-sediment exchanges, hypolimnetic mixing, and horizontal and vertical transport.
In this talk, I will present results from several studies conducted in Lake Geneva, Western Europe’s largest lake (max. depth 300 m), combining field observations, 3D numerical modelling, and particle tracking. The first part of the talk will cover the dynamics and ecological implications of wintertime coastal upwelling and interbasin exchange and upwelling, highlighting their role in deepwater renewal. The second part of the talk will present recent findings on the importance of different vertical modes of rotationally-modified standing internal waves (i.e., Kelvin and Poincaré waves) on the deepwater dynamics in Lake Geneva, highlighting the impact of seemingly negligible but ever-present weak stratification in the deep hypolimnion on the vertical structure of higher vertical-mode Poincaré waves.
- Speaker: Rafael Reiss (University of Cambridge)
- Thursday 05 June 2025, 11:30-12:30
- Venue: Open Plan Area, Institute for Energy and Environmental Flows, Madingley Rise CB3 0EZ.
- Series: Institute for Energy and Environmental Flows (IEEF); organiser: Catherine Pearson.
Thu 12 Jun 11:00: Variational Uncertainty Decomposition for In-Context Learning
As large language models (LLMs) gain popularity in conducting prediction tasks in-context, understanding the sources of uncertainty in in-context learning becomes essential to ensuring reliability. The recent hypothesis of in-context learning performing predictive Bayesian inference opens the avenue for Bayesian uncertainty estimation, particularly for decomposing uncertainty into epistemic uncertainty due to lack of in-context data and aleatoric uncertainty inherent in the in-context prediction task. However, the decomposition idea remains under-explored due to the intractability of the latent parameter posterior from the underlying Bayesian model. In this work, we introduce a variational uncertainty decomposition framework for in-context learning without explicitly sampling from the latent parameter posterior, by optimising auxiliary inputs as probes to obtain an upper bound to the aleatoric uncertainty of an LLM ’s in-context learning procedure. Through experiments on synthetic and real-world tasks, we show quantitatively and qualitatively that the decomposed uncertainties obtained from our method exhibit desirable properties of epistemic and aleatoric uncertainty.
- Speaker: Yingzhen Li (Imperial College London)
- Thursday 12 June 2025, 11:00-12:30
- Venue: Cambridge University Engineering Department, CBL Seminar room BE4-38..
- Series: Machine Learning Reading Group @ CUED; organiser: Xianda Sun.
Thu 12 Jun 11:00: Variational Uncertainty Decomposition for In-Context Learning
As large language models (LLMs) gain popularity in conducting prediction tasks in-context, understanding the sources of uncertainty in in-context learning becomes essential to ensuring reliability. The recent hypothesis of in-context learning performing predictive Bayesian inference opens the avenue for Bayesian uncertainty estimation, particularly for decomposing uncertainty into epistemic uncertainty due to lack of in-context data and aleatoric uncertainty inherent in the in-context prediction task. However, the decomposition idea remains under-explored due to the intractability of the latent parameter posterior from the underlying Bayesian model. In this work, we introduce a variational uncertainty decomposition framework for in-context learning without explicitly sampling from the latent parameter posterior, by optimising auxiliary inputs as probes to obtain an upper bound to the aleatoric uncertainty of an LLM ’s in-context learning procedure. Through experiments on synthetic and real-world tasks, we show quantitatively and qualitatively that the decomposed uncertainties obtained from our method exhibit desirable properties of epistemic and aleatoric uncertainty.
- Speaker: Yingzhen Li (Imperial College London)
- Thursday 12 June 2025, 11:00-12:30
- Venue: Cambridge University Engineering Department, CBL Seminar room BE4-38..
- Series: Machine Learning Reading Group @ CUED; organiser: Xianda Sun.
Tue 17 Jun 11:00: The ATLAS liquid argon calorimeter: an historical perspective
Abstract not available
- Speaker: Prof. Michel Lefebvre, University of Victoria
- Tuesday 17 June 2025, 11:00-12:00
- Venue: Seminar Room -- RDC D2.002 .
- Series: Cavendish HEP Seminars; organiser: Dr. Aashaq Shah.
Fri 26 Sep 08:45: CamVet Clinial Research Grants
Tom Kearns: ‘Confirming lymph node metastasis in canine mast cell tumours: A new tool in our KIT ’
Identifying genetic risk factors for intervertebral disc disease and idiopathic epilepsy in Dachshunds Bruno Lopes
Micro-computed tomography to characterise myocardial infarcts in cats with hypertrophic cardiomyopathy Jose Novo Matos
Chaired by Kate Hughes
- Speaker: Bruno Lopes, Jose Novo Matos, and Tom Kearns, Department of Veterinary Medicine
- Friday 26 September 2025, 08:45-10:00
- Venue: LT2.
- Series: Friday Morning Seminars, Dept of Veterinary Medicine; organiser: Fiona Roby.
Thu 20 Nov 13:00: Targeting the interplay between oncogenic RAS signaling and tumour immune evasion in lung cancer
Abstract not available
- Speaker: Professor Julian Downward, Francis Crick Institute
- Thursday 20 November 2025, 13:00-14:00
- Venue: CRUK CI Lecture Theatre.
- Series: Cancer Research UK Cambridge Institute (CRUK CI) Seminars in Cancer; organiser: Kate Davenport.
Tue 10 Jun 11:15: Intuitive knowledge systems for discovery
Join us for an exploration of how intuitive knowledge systems might complement current approaches in scientific discovery. Drawing from conversations during her fellowship at the Cavendish, artist Akeelah Bertram examines the acknowledged limits of current calculation systems and the role of intuition for receiving unknown phenomena. Through readings from her developing publication “Sacred Architecture,” she reflects on parallel knowledge systems, drawing from Caribbean congregational practices and embodied ways of knowing. This talk explores questions about the convergence of rigorous scientific inquiry with intuitive methodologies, considering what might emerge when different ways of knowing are held in dialogue.
- Speaker: Akeelah Bertram (Cavendish Arts Science Fellow)
- Tuesday 10 June 2025, 11:15-12:00
- Venue: Coffee area, Battcock Centre.
- Series: Hills Coffee Talks; organiser: Charles Walker.
Wed 11 Jun 13:30: A near-optimal quadratic Goldreich-Levin algorithm
In this talk I will present an efficient algorithm for a central problem in quadratic Fourier analysis, and which can be seen as a quadratic generalisation of the celebrated Goildreich-Levin algorithm. More precisely, given a bounded function f on the Boolean hypercube {0, 1}n and any ε > 0, our algorithm returns a quadratic polynomial q: {0, 1}n → {0, 1} so that the correlation of f with the function (−1)q is within an additive ε of the maximum possible correlation with a quadratic phase function. This algorithm runs in O(n3) time and makes O(n2 log n) queries to f. As a corollary, we obtain an algorithmic inverse theorem for the order-3 Gowers norm with polynomial guarantees.
Our algorithm is obtained using ideas from recent work on quantum learning theory. Its construction significantly deviates from previous approaches based on algorithmic proofs of the inverse theorem for order-3 Gowers norms (and in particular does not rely on the recent resolution of the polynomial Freiman-Ruzsa conjecture).
Based on joint work with Jop Briët.
- Speaker: Davi de Castro Silva (University of Cambridge)
- Wednesday 11 June 2025, 13:30-15:00
- Venue: MR4, CMS.
- Series: Discrete Analysis Seminar; organiser: Julia Wolf.
Mon 09 Jun 14:00: Critical norm blow-up rates for the energy supercritical nonlinear heat equation
We study the behavior of the scaling critical Lebesgue norm for blow-up solutions to the nonlinear heat equation (the Fujita equation). For the energy supercritical nonlinearity, we give estimates of the blow-up rate for the critical norm. This is based on joint work with Jin Takahashi (Institute of Science Tokyo) and Hideyuki Miura (Institute of Science Tokyo).
- Speaker: Tobias Barker (University of Bath)
- Monday 09 June 2025, 14:00-15:00
- Venue: MR13.
- Series: Partial Differential Equations seminar; organiser: Giacomo Ageno.
Thu 04 Jun 17:00: Positional encodings in LLMs
Positional encodings are essential for transformer-based language models to understand sequence order, yet their influence extends far beyond simple position tracking. This talk explores the landscape of positional encoding methods in LLMs and reveals surprising insights about how these architectural choices shape model behavior.
We begin with the fundamental challenge: why attention mechanisms require explicit positional information. We then survey the evolution of encoding strategies, from sinusoidal approaches to modern techniques like RoPE, examining their architectural implications and trade-offs.
The talk delves into how these different encoding strategies fundamentally shape model architectures and representations. We analyze the specific limitations and trade-offs of each approach, examining how positional information propagates through transformer layers and influences the learned representations.
- Speaker: Valeria Ruscio
- Thursday 04 June 2026, 17:00-17:45
- Venue: Lecture Theatre 2, Computer Laboratory, William Gates Building.
- Series: Foundation AI; organiser: Pietro Lio.