RHUL CS seminars

Venue

MCCREA 1-16 [No MS-Teams Link Available Yet]

Short Bio

Abstract

TBA

Venue

MCCREA 1-16 [No MS-Teams Link Available Yet]

Short Bio

Abstract

TBA

Venue

MCCREA 1-16 [No MS-Teams Link Available Yet]

Short Bio

Abstract

TBA

Venue

TBA [No MS-Teams Link Available Yet]

Short Bio

Abstract

TBA

Venue

MCCREA 1-16 [No MS-Teams Link Available Yet]

Short Bio

Shalini Maiti studied Information and Communication Technology at DA-IICT in India, did her masters at TU Graz, specializing in Computer Vision and Machine Learning. She is currently in the final year of her PhD at University College London and Meta AI. Shalini works with 3D vision, particularly with reconstruction, evaluation and generation of 3D objects. When she's is not huddled over academic deadlines, she spends time with good stories, shared experiences, pub quizzes and travel.

Abstract

Reconstructing and evaluating 3D content poses challenges at both the modeling and assessment stages. For non-rigid objects, 3D recovery from 2D keypoints is ill-posed due to occlusions and entanglement between viewpoint and shape variation. Classical low-rank models impose global constraints but suffer from alignment difficulties and limited expressivity. By instead constraining localized subsets of shape within high-capacity unsupervised models, it is possible to preserve flexibility while ensuring geometric consistency, yielding over 70% error reduction on S-Up3D. In parallel, the rapid growth of text-to-3D generation has exposed limitations of current evaluation metrics, which either require ground-truth supervision (e.g., PSNR) or only measure prompt fidelity (e.g., CLIP). Gen3DEval addresses this by leveraging vision-language models fine-tuned for 3D object quality assessment, enabling reference-free evaluation across text fidelity, appearance, and surface geometry. Together, these approaches advance both the generative and evaluative foundations of 3D vision, highlighting pathways toward more accurate, scalable, and human-aligned 3D understanding.

Venue

MCCREA 1-16 [MS Teams Link]

Short Bio

Arjan Cornelissen is a postdoc at Simon’s Institute, UC Berkeley. He is an expert in quantum algorithms, developing new algorithmic techniques and frameworks, and is also interested in quantum query, time and space lower bounds. Arjan received his PhD at the University of Amsterdam under the supervision of Maris Ozols, before moving on to a postdoc position at Universite de Paris Cite.

Abstract

In this talk, I will give a glimpse of an ongoing project on the development of quantum algorithms through composition of graphs. At a high level, the idea is to represent a quantum query algorithm as a graph. The graph composition framework, that this work introduces, then provides a black-box way to turn such graphs into quantum algorithms. It turns out that this framework is able to unify many existing frameworks that generate quantum query algorithms, and it provides tangible ways to make these implementations time-efficient too. If time permits, we will also take a look at several examples for which this framework can be used.

Venue

MCCREA 1-16 [MS Teams Link]

Short Bio

Mauro Vallati is a UKRI Future Leaders Fellow and ACM Distinguished Speaker on Artificial Intelligence (AI) for the UK. He is a Professor of AI at the University of Huddersfield, where he leads the Autonomous Intelligent Systems research center and the AI for Urban Traffic Control research team. Mauro has extensive experience in real-world applications of AI methods and techniques, spanning from healthcare to train dispatching. In 2014, he started working on AI applied to the field of urban traffic control, a line of research that led to numerous high-impact academic publications, patents filed in the United Kingdom, China, and the United States, as well as the deployment of the resulting techniques in urban areas of the United Kingdom.

Abstract

The current increase in urbanisation, coupled with the socio-economic motivation for increasing mobility, is pushing the transport infrastructure well beyond its capacity. Traditional urban traffic control techniques are struggling to cope with the dramatic rise of traffic, and have limited ability to react. In response, more intelligent control mechanisms are required to better monitor and exploit the available infrastructure. Despite the growing number of studies leveraging on machine learning techniques to perform traffic control tasks, the good old model-based AI is gaining traction, thanks to its ability to provide approaches that can smoothly deal with unusual and unexpected conditions. In this talk we will focus on the application of AI planning to urban traffic control. First, we will look into how urban traffic control is currently performed and what are the main challenges faced. Then, we will present how AI planning techniques have been used to improve common practices, and to provide useful tools for traffic engineers, domain experts, and practitioners.

Venue

TOLANSKY 125 [MS Teams Link]

Short Bio

Rafael Izbicki is a faculty member in the Department of Statistics at the Federal University of São Carlos, Brazil, and has been a CNPq research fellow since 2017. He earned his BSc (2009) and MSc (2010) in from the University of São Paulo and his PhD (2014) from Carnegie Mellon University. Rafael leads the Statistical Machine Learning Lab, with over 80 peer-reviewed publications and is the author of the book Machine Learning Beyond Point Predictions: Uncertainty Quantification. His research covers fundamental aspects of statistics and machine learning, with applications in biology, medicine, epidemiology, and cosmology.

Abstract

Building confidence sets that reach their nominal level is especially hard in complex models and when the observed sample size is small, as in many likelihood-free inference tasks. We introduce a simulation-based calibration scheme that constructs local confidence sets using only samples drawn from the data-generating model. Adapting insights from conformal prediction to parameter inference, the method forms data-adaptive partitions that deliver finite-sample local coverage and achieve conditional coverage as the number of simulations grows. It handles nuisance parameters with no substantial computational cost and computes uncertainty estimates that flag when more simulations are needed. Experiments on models with and without tractable likelihoods show that the resulting sets are better calibrated than existing alternatives. The procedure provides a practical tool for reliable inference in modern statistical applications.

Venue

MOORE-0-16 [MS Teams Link]

Short Bio

I am a PhD student at the University of Waterloo supervised by Professor Ming Li and advised by Professor Marcus Hutter. My research is on algorithmic probability (Kolmogorov complexity, Solomonoff induction), particularly as it applies to sequential decision theory. More generally, I am interested in applying ideas from theoretical computer science to understanding (artificial) intelligence, with a long-term goal of gaining traction on the mathematical foundations of superintelligent agents and the alignment problem. Previously, I earned an M.S. in mathematics at the University of Minnesota, and at various times worked on robotics and computer vision.

Abstract

It is a longstanding challenge in artificial intelligence (and for that matter, philosophy) to build agents that understand themselves as part of a larger and computationally more powerful environment. Algorithmic probability provides tools to define versatile learners and agents that represent their knowledge in the form of programs, but such agents cannot be represented as programs and therefore cannot meaningfully reflect. There have been many attempts to remedy this problem, the most elegant relying on “quining” (Kleene's second recursion theorem). Another approach is imprecise probability, which arises surprisingly naturally from certain expected utilities of these agents. I will explain why this problem is important, tie together existing and novel approaches, and explain why all of them fall short.

Venue

BEDFORD-0-07 [No MS-Teams Link Available Yet]

Short Bio

I am Lecturer in Econometrics at the University of Surrey's School of Economics. I hold a Ph.D in Economics from Universitat Pompeu Fabra and an MSc in Data Science from the Barcelona School of Economics. My main areas of interest are time series forecasting, financial econometrics, machine learning, empirical finance and macroeconomics.

Abstract

This paper develops the concept of Growth and Inflation at Risk frontier (GIaR). This is a bivariate generalisation of the concepts of Growth-at-Risk (GaR) and Inflation-at-Risk (IaR). We propose a novel approach to identify and estimate GIaR and provide uniformly valid upper and lower confidence bands. We first apply our procedure to predict the conditional probability of stagflation. Second, we compute worst-case scenarios for a policy maker who is concerned about the joint tail risk of low growth and high inflation. Our empirical findings show that a tightening of financial conditions increases joint tail risks to both growth and inflation.

Venue

BEDFORD-0-07 [MS Teams Link]

Short Bio

Laura is a Reader in Computing and the Head of the Programming Languages and Systems (PLAS) group at the University of Kent. She is interested in theories and tools to develop safe distributed systems and, in particular, static and dynamic verification of message-passing concurrency using session types. Her most recent research focuses on time-sensitive systems and reliable systems. She is PI at Kent of EPSRC project 'Session Types for Reliable Distributed Systems (STARDUST)'.

Abstract

Session types are a formalism for modelling structured communications in concurrent systems and can be seamlessly implemented into programming languages. Session subtyping answers the question of whether a program in a concurrent system can be safely substituted for another, when their communication behaviours are described by session types. Asynchronous session subtyping is undecidable in general, hence the interest in devising sound (though incomplete) subtyping algorithms. State-of-the-art methods rely on a data structure known as input trees. I will present an alternative approach that replaces input trees with sets of traces, enabling the application of abstract interpretation techniques to this problem. This trace-based approach introduces a novel way to relax (enlarge) trace sets while maintaining observable subtyping properties. Crucially, these relaxations can be finitely represented, even when the corresponding input trees are arbitrarily large. This method can be instantiated using regular expressions, enabling us to mechanically prove subtyping for communication patterns previously beyond reach.

Venue

BEDFORD-0-07 [MS Teams Link]

Short Bio

Vangelis Markakis is a Professor in the Department of Informatics at the Athens University of Economics and Business. He obtained his PhD in Computer Science at the Georgia Institute of Technology (Georgia Tech, 2005), and he has also held appointments as a postdoctoral researcher at the University of Toronto (2005-2006), and at CWI, Amsterdam (2007-2008), and as a visiting scholar at Stanford University (2014). Since 2022, he has also been a collaborating member of the Input Output Global (IOG) research group on game-theoretic aspects of blockchain protocols. His research interests lie at the intersection of game theory and theoretical computer science, with an emphasis on optimization, mechanism design, fair division, and voting theory.

Abstract

This talk focuses on game-theoretic aspects that arise in the design of reward schemes in blockchain protocols. The first part concerns a model for pool formation in Proof of Stake protocols. In such systems, stakeholders can form pools as a means of obtaining regular rewards from participation in ledger maintenance and block production, with the power of each pool being dependent on its collective stake. The question of interest here is the design of reward schemes that suitably split earned profits among pool members. With this in mind, we initiate a study of the well known Shapley value scheme into the context of blockchains. We provide comparisons between the Shapley mechanism and the more standard proportional scheme, in terms of attained decentralization (i.e., number of pools formed at equilibrium), and in terms of susceptibility to Sybil attacks, i.e., the strategic splitting of a players' stake with the intention of participating in multiple pools for increased profit. The second part of the talk is motivated by a different application scenario and investigates the impact of reward schemes on blockchain governance. We introduce a model of elections where there is a ground truth (i.e., a ``correct" outcome), and where stakeholders can only choose to delegate their voting power to a set of delegation representatives (DReps). As a way to motivate the representatives to exert effort and discover the ground truth, a reward scheme can be used based on the total delegated stake attracted by each DRep. We analyze different schemes in this context with respect to their equilibrium properties, given an available monetary budget. Our findings provide insights into the design of effective reward mechanisms and optimal committee structures (i.e., how many DReps are enough) in such systems with the goal of maximizing the probability for selecting the correct outcome.

Venue

BEDFORD-0-07 [MS Teams Link]

Short Bio

David Tena-Cucala is a Lecturer in Computer Science at Royal Holloway, University of London. He has a BSc in Mathematics and Physics, and an MPhil In Philosophy, specialising in Philosophy of Physics and science. He graduated with a D.Phil (PhD) in Computer Science from University of Oxford, where he wrote a thesis on automated reasoning algorithms for expressive Description Logics. His current research interests lie in the intersection of Knowledge Representation and Machine Learning interpretability.

Abstract

Graph Neural Networks (GNNs) are a family of Machine Learning (ML) models that can be applied to graph-shaped data. They have been highly successful in data analytics applications such as drug discovery, product recommendation, and fraud detection, among many others. GNNs, however, lack some of the features enjoyed by classic Knowledge Representation (KR) methods for data analytics, such as interpretability, verifiability, or explainability. In this talk, I will show how we can bridge the gap between GNN and KR methods. In particular, I will show how to extract KR theories from certain classes of GNNs with theoretical guarantees, so that both the KR theory and the GNN are provably equivalent. I will then show applications of this approach to GNN interpretability, verification, and explainability.

Venue

BEDFORD-0-07 [MS Teams Link]

Short Bio

Dr. Jialin Yu is a postdoctoral researcher in the Department of Statistical Science at UCL. His research lies at the intersection of machine learning, representation learning, and causal inference, with a focus on developing robust and generalizable models for real-world applications. His work leverages principled statistical methods, including Bayesian inference and causal reasoning, to enhance AI systems' ability to generalize beyond observed data.

Abstract

In this talk, I will start with a general introduction on the hierarchy of machine learning models (seeing, knowing and doing machines). Then move to introduce one of the attempt to building a certain type of doing machine, which allow us to answer the following question: "how do we generalize from past experiments to novel compositional conditions?". One of the goals of causal inference is to generalize from past experiments and observational data to novel conditions. While it is in principle possible to eventually learn a mapping from a novel experimental condition to an outcome of interest, provided a sufficient variety of experiments is available in the training data, coping with a large combinatorial space of possible interventions is hard. Under a typical sparse experimental design, this mapping is ill-posed without relying on heavy regularization or prior distributions. Such assumptions may or may not be reliable, and can be hard to defend or test. In this talk, we take a close look at how to warrant a leap from past experiments to novel conditions based on minimal assumptions about the factorization of the distribution of the manipulated system, communicated in the well-understood language of factor graph models. A postulated interventional factor model (IFM) may not always be informative, but it conveniently abstracts away a need for explicitly modeling unmeasured confounding and feedback mechanisms, leading to directly testable claims. Given an IFM and datasets from a collection of experimental regimes, we derive conditions for identifiability of the expected outcomes of new regimes never observed in these training data. We implement our framework using several efficient algorithms, and apply them on a range of semi-synthetic experiments.

Venue

BEDFORD-0-07 [MS Teams Link]

Short Bio

Dr. Michail Fasoulakis is an (Applied) Mathematician and (Theoretical) Computer Scientist, with a background and interest in Computer Engineering and (Computational) Economics/Game theory/Operations Research, and currently a Lecturer in Theoretical Computer Science at Royal Holloway, University of London and an Affiliated/Corresponding Scientist of Abroad (Honorary title) at the Institute of Computer Science of FORTH. He holds a B.Sc. in Computer Science, a second B.Sc. in Mathematics, and a M.Sc. in Communications and Signal Processing, all from the University of Crete. Furthermore, he holds a M.Sc. in Computation and Game theory from the University of Liverpool, UK. He received the Ph.D. in (Theoretical) Computer Science from the University of Warwick UK, in 2017, working in the research area of algorithmic game theory, where he was a member of the Centre for Discrete Mathematics and its Applications. After his PhD studies he had research fellow/postdoc positions in various academic places such as Athens University of Economics and Business (AUEB), Center for Mathematics and Computer Science in Netherlands - Centrum Wiskunde & Informatica/CWI (as an ERCIM "Alain Bensoussan" fellow), Foundation for Research and Technology–Hellas (FORTH) as a Stavros Niarchos Foundation (SNF)-FORTH fellow for a period, School of Electrical and Computer Engineering of the National Technical University of Athens, and University of Crete. He has also been a visiting researcher at the University of Liverpool, UK, and the University of Warwick, UK. His research interests include the (intersections of) areas of Algorithms, Game theory, Optimization, Mathematics of Information (Information theory), and their (theoretical) applications, especially in Operations Research, Economics, AI/ML, Communications, and Signal Processing. His work has been published in high quality journals and conferences' proceedings such as ACM Transactions on Algorithms, Algorithmica, IEEE Transactions on Neural Networks and Learning Systems, SIAM Journal on Computing, AAAI, AISTATS, ESA, and SODA, with notable contributions, with his coauthors, the current state of art polynomial-time algorithms for computing approximate (well-supported) Nash equilibria in bimatrix games, giving the best approximation bounds at the moment of the publications. During his studies and career, he has been awarded with a number of research grants/fellowships-scholarships/awards.

Abstract

By the seminal paper of John von Neumann on the Minimax Theorem for two-player zero-sum games, this class of games became one of the most fundamental class in game theory. It is well-known that the computation of a pair of optimal solutions, Nash equilibrium, in this class can be done in polynomial-time by a number of different algorithms such as algorithms for Linear programming, fist-order methods and learning algorithms. However, many of these algorithms are centralized and/or converge to an optimal solution on average. On the other hand, recent applications of zero-sum games in machine learning, particularly Generative Adversarial Networks (GANS), have created the need of new algorithms with better properties such as last-iterate convergence to an optimal solution. In this presentation, we present our two recent algorithms for computing a Nash equilibrium in these games: one gradient-based algorithm and one learning algorithm that improve the state of the art from different perspectives.

Venue

BOURNE-LT2 [MS Teams Link]

Short Bio

Christian Igel is a professor at DIKU, the Department of Computer Science at the University of Copenhagen. He studied Computer Science at the Technical University of Dortmund, Germany. In 2002, he received his Doctoral degree from the Faculty of Technology, Bielefeld University, Germany, and in 2010 his Habilitation degree from the Department of Electrical Engineering and Information Sciences, Ruhr-University Bochum, Germany. From 2003 to 2010, he was a Juniorprofessor for Optimization of Adaptive Systems at the Institut für Neuroinformatik, Ruhr-University Bochum. In October 2010, he was appointed professor with special duties in machine learning at DIKU. Since December 2014 he has been a full professor at DIKU. His main research area is Machine Learning. Currently, he is particularly interested in support vector machines and other kernel-based methods, evolution strategies for single- and multi-objective optimization and reinforcement learning, PAC-Bayesian analysis of ensemble methods, deep neural networks and stochastic neural networks, and applications of machine learning that help achieve the sustainable development goals.

Abstract

Tree-based ecosystems play an important role for carbon sequestration and biodiversity as well as for timber and food production. We need a better characterization of woody resources at a global scale to understand how these ecosystems are affected by climate change and human management. Recent advances in remote sensing and machine learning (ML) based computer vision make this possible. Deep learning applied to high-resolution satellite imagery allows, for example, for large-scale mapping of individual trees. The biomass of each tree, and thereby its carbon content, can then be estimated from the crown size using allometric equations learned from data. The functional relation is assumed to be non-decreasing, and incorporating this constraint into ML-based modelling increase the scientific plausibility and acts as a strong regularizer. We highlight an application of tree carbon stock quantification in Rwanda, where it helps developing pathways to reach the country’s sustainable development goals. The talk closes with a presentation of recently released open source image data sets for training ML models for different environmental monitoring tasks.

Venue

BOURNE-LT2 [MS Teams Link]

Short Bio

Sophie Klumper is a fourth-year PhD student in the Networks and Optimization Group at Centrum Wiskunde & Informatica (CWI), where she is advised by Guido Schäfer. Her primary research interests are algorithmic game theory and approximation algorithms.

Abstract

The realm of algorithms with predictions has led to the development of several new algorithms that leverage (potentially erroneous) predictions to enhance their performance guarantees. The challenge here is to devise algorithms that achieve optimal approximation guarantees as the prediction quality varies from perfect (consistency) to imperfect (robustness). This framework is particularly appealing in mechanism design contexts, where predictions might convey private information about the agents, motivating our research. Our goal is to design strategyproof mechanisms that leverage predictions to achieve improved approximation guarantees for several variants of the Generalized Assignment Problem (GAP) in the private graph model. In this model, first introduced by Dughmi and Ghosh (2010), the set of resources that an agent is compatible with is private information, and assigning an agent to an incompatible resource generates zero value. For the Bipartite Matching Problem (BMP), we give a deterministic group-strategyproof (GSP) mechanism that is (1+1/γ)-consistent and (1+γ)-robust, where γ ≥ 1 is some confidence parameter. We also prove that this is best possible. Remarkably, our mechanism draws inspiration from the renowned Gale-Shapley algorithm, incorporating predictions as a crucial element. Additionally, we consider multiple variants of GAP for which we provide deterministic or randomized mechanisms that are (universally) GSP. All our mechanisms also provide more fine-grained approximation guarantees that smoothly interpolate between the consistency and robustness guarantees, depending on some natural error measure of the prediction.

Venue

BOURNE-LT2 [MS Teams Link]

Short Bio

Antonio is a Research Associate in the Department of Computing at Imperial College London, where he also completed a PhD titled “Gradual Evaluation in Argumentation Frameworks: Methods, Properties and Applications” under the supervision of Prof. Francesca Toni. His main research interests lie in the deployment of argumentative technologies to applications, particularly in explainable AI, e.g. for explaining the outputs of neural networks, recommender systems or Bayesian classifiers, but also in other settings such as e- engineering, medicine, democracy and judgemental forecasting.

Abstract

As data-driven AI models achieve unprecedented feats across previously unthinkable tasks, the diminishing levels of interpretability of their increasingly complex architectures can often be sidelined in place of performance. If we are to comprehend and trust these AI models as they advance, it is clear that symbolic methods, given their unparalleled strengths in knowledge representation and reasoning, can play an important role in explaining AI models. In this talk, I discuss some of the ways in which one branch of such methods, computational argumentation, given its human-like nature, can be used to tackle this problem. I first outline a general paradigm for this area of explainable AI, before detailing a prominent methodology therein which we have pioneered. I then illustrate how this approach has been put into practice with diverse AI models and types of explanations, before looking ahead to challenges, future work and the outlook in this field.

Venue

BOURNE-LT2 [MS Teams Link]

Short Bio

Vladimir Vovk is a Professor at Royal Holloway, and co-director of the CRML. He is the co-inventor of Conformal Prediction and game-theoretic probability.

Abstract

I will start from a brief review of e-values, which are becoming popular as an alternative to p-values in statistical hypothesis testing. Conformal e-prediction is the counterpart of conformal prediction for e-values. Conformal e-prediction is conceptually simpler and had been developed in the 1990s as precursor of conformal prediction. When conformal prediction emerged as result of replacing e-values by p-values, it seemed to have important advantages over conformal e-prediction without obvious disadvantages. The work reported in this talk re-examines relations between conformal prediction and conformal e-prediction systematically from a modern perspective. Conformal e-prediction has advantages of its own, such as the ease of designing conditional conformal e-predictors and the guaranteed validity of cross-conformal e-predictors (whereas for cross-conformal predictors validity is only an empirical fact and can be broken with excessive randomization). Even where conformal prediction has clear advantages, conformal e-prediction can often emulate those advantages, more or less successfully.

Venue

BOURNE-LT2 [MS Teams Link]

Short Bio

Agnieszka Mensfelt is a post-doctoral research assistant at Royal Holloway, University of London. She obtained a master's in Cognitive Science from Adam Mickiewicz University and a master's and PhD in Computer Science from Poznan University of Technology. Prior to joining Royal Holloway, she worked as a research assistant at Poznan University of Technology. Her research interests include agent-based modelling, evolutionary optimization, and game theory. She is one of the co-developers of the Framsticks simulation environment.

Abstract

Artificial life is a fascinating research field that draws from biology, chemistry, computer science, and the social sciences to investigate the fundamental principles of life. This presentation will focus on the Framsticks environment – a versatile software that allows users to model, simulate, and optimize artificial life forms. The applications include simulating multi-agent systems, performing evolutionary optimization, and modelling agent control, adaptation, communication, and interaction. The talk will provide an overview of the goals and applications of artificial life, followed by an introduction to Framsticks. It will also discuss selected use cases of the platform.

Venue

Bedford 0-07 [MS Teams Link]

Short Bio

Harris Papadopoulos is an Associate Professor at the Department of Electrical Engineering, Computer Engineering and Informatics of Frederick University, where he is heading the Computational Intelligence (COIN) Laboratory. He is also a Visiting Professor at the Computer Science Department of Royal Holloway, University of London. His research primarily focuses on the development of advanced Machine Learning techniques that provide probabilistically valid uncertainty quantification for each prediction under minimal assumptions, and on their application to a diverse array of challenges in various scientific fields. He has published more than 75 research papers in prestigious international refereed journals and conferences. He acted as co-editor of the book “Measures of Complexity: Festschrift in Honor of Alexey Chervonenkis” published by Springer in 2015, which was dedicated to one of the most influential figures in the field of Machine Learning. He has been involved in several national and EU funded research projects, frequently taking the lead as the coordinator and initiator, with a total budget of more than 6M euro. Direct funding to his lab in the past five (5) years exceeds 1.5M euro. He has also secured an industrial grant (of around 400,000 euro) for the application of Machine Learning techniques to large volumes of data, from one of the biggest international consulting companies for complicated assets. He served as an Associate Editor of the Evolving Systems Journal (Springer) from 2015 to 2018 and as guest editor and reviewer for numerous esteemed scientific journals. Additionally, he served as one of the main organizers of several international conferences and workshops (General Chair of COPA 2023, AIAI 2013, Organizing Chair of AIAI 2010, PC Chair of eight conferences and Chair of nine workshops).

Abstract

Conformal Prediction (CP), a framework pioneered at Royal Holloway’s Centre for Reliable Machine Learning, has marked a significant stride towards more reliable machine learning models. CP stands out as a versatile framework capable of offering prediction regions with valid coverage guarantees under minimal assumptions. Its unique strength lies in providing probabilistically valid guarantees, which are of paramount importance for complex tasks across various domains where decision certainty is crucial. This talk will introduce the CP framework and discuss the recent advancements made by the COmputational INtelligence (COIN) research laboratory at Frederick University in Cyprus. Maintaining strong collaborative ties with Royal Holloway, COIN has significantly contributed to the ongoing development and application of CP across diverse fields spanning from biomedicine and engineering to entertainment and agriculture. The aim of this talk is to showcase the foundational and applied research contributions of COIN, fostering dialogue and exploring potential collaborations with attendees interested in advancing the field of reliable machine learning.

Venue

MOORE 0-16 [MS Teams Link]

Short Bio

George Osipov is originally from Tbilisi, Georgia, where he did his bachelor’s in computer science and studied towards a master’s degree in mathematics, with an Erasmus year in Uppsala, Sweden. In June 2024 he defended his PhD thesis in theoretical computer science titled “On Infinite-Domain CSPs Parameterized by Solution Cost” at Linköping University, Sweden, under supervision of Peter Jonsson and Victor Lagerkvist. He received a VR international postdoc grant to continue as a postdoc at Linköping University with plans to visit Oxford University for a year in 2025 and Royal Holloway for a year in 2026. He is working mostly on constraint satisfaction and parameterized complexity.

Abstract

Based on joint work with Marcin Pilipczuk and Magnus Wahlström to appear at ESA’24. The input to the Minimum-Cost Constraint Satisfaction Problem (MinCSP) over the Point Algebra is a set of variables, a collection of constraints of the form x<y, x=y, x≤y or x≠y, and a budget k. The goal is to check whether one can assign rational values to the variables while breaking at most k constraints. By restricting the set of allowed constraint types, this problem specializes to several prominent graph problems such as Directed Feedback Arc Set which is MinCSP(<), its subset variant which is MinCSP(<,≤) and Edge Multicut which is MinCSP(=,≠). All three are famously NP-hard but fixed-parameter tractable (FPT), i.e. they admit algorithms with running times of the form f(k) poly(n), where f is a function that depends solely on the budget k, while poly(n) is some polynomial in the input size. Another special case is MinCSP(≤, ≠) also known as Directed Symmetric Multicut. Fixed-parameter tractability of the latter was an open problem prior to this work. We obtain a complete classification of parameterized complexity of MinCSP(Γ) for all subsets Γ of {<, ≤, ≠, =}. Towards this, we construct an algorithm that solves MinCSP(<,=,≠) - a common generalization of Directed Feedback Arc Set and Edge Multicut. We complement this with a lower bound showing that Directed Symmetric Multicut is W[1]-hard. In the talk I will describe both the ideas behind the algorithm and the lower bound construction. I will also talk about the classification project for Temporal MinCSP - a vast generalization of MinCSP over Point Algebra where all relations first-order definable using < are allowed in the constraints.

Venue

Bedford 0-07 [MS Teams Link]

Short Bio

Michelle specialized in the study of graph structures during her bachelor's degree in mathematics at TU Berlin. During her master's in computer science, she further explored graph theory and was introduced to social choice theory, culminating in her master's thesis in 2022. Currently, she is a PhD student at HPI in Potsdam, focusing her research on temporal graph structures and voting methods.

Abstract

River is a single-winner, Condorcet-consistent voting method that is based on pairwise margins and can be seen as a simplified variation of Tideman's Ranked Pairs method. Like Ranked Pairs and Schulze's Beat Path, River is a refinement of the Split Cycle method and shares many desirable properties with these. Unlike the other three methods, River satisfies a strong form of resistance to agenda-manipulation that is known as independence of Pareto-dominated alternatives. We will look at axiomatic properties of River and analyse our experimental results to establish that River provides a compromise between Ranked Pairs and Beat Path.

Venue

Bedford 0-07 [MS Teams Link]

Short Bio

Sanjukta Roy is a lecturer in Theoretical Computer Science at University of Leeds. Previously, she was a postdoctoral researcher at TU Wien, Vienna, hosted by Jiehua Chen, CVUT Prague, hosted by Dusan Knop, and Penn State University, hosted by Hadi Hosseini. She received her Ph.D. at IMSc, Chennai, in Jan 2021, where her advisor was Prof. Saket Saurabh. Her thesis is titled ``Select, Allocate, and Manipulate via Multivariate Analysis''. Her areas of specialization are Algorithmic Social Choice, Algorithms, Parameterized Algorithms, and Algorithmic Game Theory. Her research interests include Stable Matching, Fair Allocation, Hedonic games, Voting Theory, and strategic behavior of agents with preferences.

Abstract

"Social Choice or Societal Decision Making is concerned with how to aggregate individual opinions or interests. This is the common goal of a wide variety of problems including political elections, evaluation processes for job applicants or products, and recommendation systems in marketing. Two central challenges that must be taken into account from the perspective of a computer scientist are: (1) How time-efficient is it to find a good decision? (2) Is it possible to manipulate the decision? I will discuss these two questions in the context of voting theory, namely, we will discuss The Committee Selection problem and the Gerrymandering problem. Towards this, we will solve a directed cut problem, a variant of a longest path problem, and use a polynomial method to solve a problem related to graph partitioning."

Venue

Bedford 0-07 [MS Teams Link]

Short Bio

Anna Calissano is a Chapman Fellow at Imperial College London. She received her PhD at Politecnico di Milano in collaboration with Danmarks Tekniske Universitet (DTU) in Denmark. Her research lays at the intersection between geometry and statistics and focuses on defining statistical methods for the analysis of sets of graphs and networks. Her methods have been applied to landscape designing, public transport modelling, and the analysis of medical images.

Abstract

Sets of graphs (or networks) arise in many different fields, from medicine to finance, from sport to the social sciences. The analysis of unlabeled graphs or networks is far from trivial due to the highly non-Euclidean nature of such data. We describe Graph Space as a possible geometric embedding for a set of unlabeled graphs, i.e. graphs with no node correspondence across observations. Graph Space is a quotient space, but it is not a manifold, requiring the definition of statistical methods beyond the tangent space approach. We introduce the Align All and Compute algorithm and use it for both estimating generalized geodesic principal components and generalized geodesic regression models, showing how to interpolate between unlabeled graphs. We demonstrate the flexibility of the framework on both simulated data, public transport system data and Fifa 2018 player passing network data.

Venue

Bedford 0-07 [MS Teams Link]

Short Bio

Tiger-Lily recently graduated from an MSci in CompSci at RHUL. Now she is a Teaching Fellow and PhD student at RHUL under the supervision of Eduard Eiben and Argyrios Deligkas in the department. Her main interests lie in computational social choice and algorithms and complexity, specifically parametrized complexity.

Abstract

In this talk we will see the problem of finding fair allocations -- EF1 and EFX -- of indivisible goods with set-restrictions, which we call orientations. In an orientation every agent gets items from their own predetermined set. For EF1, we show a surprisingly positive result. They always exist! For EFX, we focus on the recently proposed graph instances, where every agent corresponds to a vertex on a graph and their allowed set of items consists of the edges incident to their vertex. However, it was shown that finding an EFX orientation is NP-complete in general. We prove that it remains intractable in some very restricted cases. We essentially match these strong negative results with a fixed parameter algorithm that is virtually the best someone could hope for.

Venue

Founder West 101 [MS Teams Link]

Short Bio

Bipin Rajendran is a Professor of Intelligent Computing Systems and EPSRC Fellow at King's College London (KCL). He received a B. Tech degree from I.I.T. Kharagpur in 2000, and MS and PhD degrees in Electrical Engineering from Stanford University in 2003 and 2006, respectively. He was a Master Inventor and Research Staff Member at IBM T. J. Watson Research Center in New York from 2006-'12 and has held faculty positions in India and the US. His research focuses on building algorithms, devices, and systems for brain-inspired computing. He has co-authored over 95 papers in peer-reviewed journals and conferences, one monograph, one edited book, and 59 issued U.S. patents. He is a recipient of the IBM Faculty Award (2019), IBM Research Division Award (2012), and IBM Technical Accomplishment Award (2010). He was elected a senior member of the US National Academy of Inventors in 2019.

Abstract

The complexity of AI models have been increasing exponentially, leading to exploding energy costs for training and inference. Furthermore, they often provide unreliable results and predictions, as they are unable to account for uncertainty in models or data. In this talk, I will discuss our recent results on designing efficient and trustworthy AI hardware based on nanoscale memristive devices. I will discuss how hardware-aware training of AI models can be used to mitigate the effect of device noise to implement in-memory computing architectures and achieve over 100-fold improvement in inference efficiency compared to CMOS designs. I will also describe how we can leverage nanoscale device stochasticity to implement Bayesian AI models that are able to provide quantifiable metrics of uncertainty associated with their decisions.

Venue

Bedford 0-07 [No MS-Teams Link Available Yet]

Short Bio

Milind Chabbi is a Senior Staff Researcher in the Programming Systems Team at Uber, USA. He leads initiatives across Uber in the areas of compiler optimizations, high-performance parallel computing, synchronization techniques, and performance analysis tools to make large, complex computing systems reliable and efficient. Milind is also adept at code analysis and automation tools that improve developer velocity. Milind earned his Ph.D. in computer science from Rice University.

Abstract

Go is the language of choice at Uber for developing microservices and infrastructure tools. Go brings concurrent programming to the masses. The presence of shared memory alongside message passing makes Go a unique programming language. Writing correct and efficient parallel programs is hard; Go does not guarantee the correctness or efficiency of parallel programs. In this tech talk, I will discuss the landscape of concurrency bugs in Go and how they impact Uber's production software systems. Using data races as an example of shared-memory concurrency bugs and deadlocks as an example of message-passing bugs, I will discuss the tools we have designed and deployed over the past several years to detect and mitigate these bugs at Uber. The insights gained from close inspection of concurrency bugs in Go reveal the complex interplay between language design and concurrency bugs.

Venue

Bedford 0-07 [MS Teams Link]

Short Bio

Marcos Matabuena, a Spanish mathematician and current postdoctoral researcher at Harvard University, specializes in the development of mathematical models applied to digital medicine. His research focusejs on key areas such as uncertainty quantification, survival analysis, and causal inference, essential aspects for the accurate prediction of clinical events and the optimization of treatments in complex diseases. In the field of diabetes, for example, his models are fundamental for interpreting data from continuous glucose monitors. Additionally, his studies on aging-related decline, through the analysis of physical activity patterns recorded by wearable devices, are of great importance given the aging of modern populations and changing demographic scenarios. One of his most notable contributions is the development of the first mathematical framework for uncertainty quantification in regression models in metric spaces. Marcos has also introduced the innovative concept of 'glucodensity,' an advanced functional representation for interpreting continuous glucose monitoring data and other data collected through wearable devices, surpassing many aspects of traditional clinical biomarkers. Another significant contribution is the development of the first biclustering algorithm in RKHS spaces for complex data analysis

Abstract

Uncertainty quantification is crucial in the era of data-driven systems, providing confidence and accuracy in decision-making, particularly in the field of biomedicine. In this area, clinical decisions often rely on the uncertainty associated with the outcomes of various predictive algorithms. In the context of digital medicine and the evolution of biological measurement systems, clinical data are represented as statistical entities in metric spaces, such as density functions or complex networks, offering a deeper understanding of biological systems. In this presentation, we focus on introducing novel conformal inference algorithms designed for metric spaces, developed by the authors, and demonstrate their application in digital medicine with the high-resolution data mentioned. These novel methods provide robust guarantees, including optimal convergence rates, which are particularly useful when both the predictor and the response are in separable Hilbert spaces. Adaptations of these algorithms to challenges, such as incomplete information (for example, missing data), will also be explored. A significant case study is the development of algorithms for personalized data imputation. By applying this technique and integrating it with digital biomarkers, we have managed to increase the accuracy of predicting the development of diabetes by more than 20%, compared to traditional methods.

Venue

Bedford 0-07 [MS Teams Link]

Short Bio

Dr. Argyrios Deligkas is a Senior Lecturer at the Computer Science Department. His research focuses on economics and computation and more specifically in equilibrium computation, algorithmic game theory, computational social choice, and fair division. He has published over 40 papers in top peer-reviewed conferences in AI, and theoretical computer science, and 16 journals in top peer-reviewed journals.

Abstract

We study the problem of computing approximate market equilibria in Fisher markets with separable piecewise-linear concave (SPLC) utility functions. In this setting, the problem was only known to be PPAD-complete for inverse-polynomial approximations. We strengthen this result by showing PPAD-hardness for constant approximations. This means that the problem does not admit a polynomial time approximation scheme (PTAS) unless PPAD=P. In fact, we prove that computing any approximation better than 1/11 is PPAD-complete. As a direct byproduct of our main result, we get the same inapproximability bound for Arrow-Debreu exchange markets with SPLC utility functions. This is joint work with John Fearnley (Liverpool), Alexandros Hollender (Oxford), Themistoklis Melissourgos (Essex).

Venue

Bedford 0-07 [MS Teams Link]

Short Bio

Sudipta Chattopadhyay received the Ph.D. degree in computer science from the National University of Singapore, in 2013. He is an Assistant Professor with the Information Systems Technology and Design Pillar, Singapore University of Technology and Design, Singapore. His general research interests lie in the broad area of cyber security including but not limited to security for AI, Wireless Technologies, and Internet of Things (IoTs). Together with his student, he discovered SweynTooth, BrakTooth and 5Ghoul, families of Bluetooth and 5G NR vulnerabilities that affect billions of devices worldwide. His research has been featured in WIRED, PC Magazine, Hacker News, among others. His discovery has also generated cyber security alerts from government regulatory agencies including CSA (Singapore), DHS, and FDA.

Abstract

In this talk, I will provide an overview of our past few years' work on Wireless Fuzzing. In particular, I will outline some key challenges that we faced for automatically testing COTS wireless devices over-the-air and how challenges were overcome to develop our security testing technology. I will discuss how our technology had evolved from a generational approach to what we call now man-in-the-middle fuzzing and how such a fuzzing method found more than 70+ security flaws in wireless devices implementing diverse range of protocols e.g., Wi-Fi, BLE, Bluetooth BR/EDR, Zigbee, CoAP and more recently 5G NR.

Venue

Bedford 0-07 [MS Teams Link]

Short Bio

Abstract

Recent advances in machine learning (ML) have demonstrated impressive performance on complex tasks such as human-level image understanding and natural language processing. These ML models rely on artificial neural networks that, like biological brains, use billions of neurons and synapses to process complex stimuli. Unlike biological brains, however, these neural networks consume untold amounts of energy, with a single training session often exceeding the energy and carbon footprint of a car over its lifetime. If the current rate of growth continues, ML models could overtake the transportation sector in the global energy balance in 10-20 years, posing another major threat to mitigating climate collapse. In this talk, I will highlight the mechanisms that enable the amazing energy efficiency of biological brains. Based on these findings, I will present new approaches to significantly reduce the energy footprint using hybrid ML/bio-inspired models.

Venue

Bedford 0-07 [MS Teams Link]

Short Bio

Abstract

Recent years have seen remarkable progress in the capabilities of large deep learning models, especially in language processing. Large language models such as GPT-3 derive a lot of their utility from being zero or few-shot learners, which are instances of a phenomenon labelled in-context learning -- the ability to perform a diverse range of tasks simply based on demonstration examples provided as input context, without any gradient updates to the model parameters. This emergent capability is poorly understood and has intrigued the ML research community. In this talk, I will provide an overview of in-context learning and survey the current state of understanding regarding the emergence of these new abilities in large language models. Then, I will examine the extent to which in-context learning needs large-scale models and the contribution of model architectures versus training approaches. Finally, I will analyse and relate this phenomenon to meta-learning approaches with recurrent architectures that predate large language models.

Venue

Bedford 0-07 [MS Teams Link]

Short Bio

Abstract

The model of population protocols provides a universal platform to study distributed processes driven by pairwise interactions of anonymous agents. Pairs of agents are drawn by the random scheduler uniformly at random from a large population of size n. Each interacting pair is formed of the initiator and the responder. On the conclusion of an interaction the states of the two agents are modified according to the predefined transition function, i.e., the set of rules underpinning the considered process. The state space S of agents is fixed and the size of the population is not known, i.e., not hard-coded in the transition function. The sequential time of stabilisation of a population protocol refers to the number of interactions required to reach one of the final configurations. More recently, the focus is on the parallel time (or simply time) defined as the sequential time divided by n, where a given protocol is efficient if it stabilises in parallel time O(poly log n). Finally, we say that a population protocol is stable if it stabilises with the correct solution with probability 1. Among several computational deficiencies of (also stable) population protocols are: depleting fraction of meaningful interactions closing in on the final stabilisation (suppressing parallel efficiency), computation power of constant-space population protocols limited to semi-linear predicates in Presburger arithmetic (reflecting on the time-space trade offs), and indefinite computation (impacting multi-stage protocols). With these deficiencies in mind, we propose a new selective variant of population protocols by imposing an elementary structure on the state space, together with a conditional choice of the responder during random interacting pair selection. We first observe that selective protocols provide a natural mechanism for deterministic emptiness test (also known as zero test). It is known that such test enables efficient simulation of O(log n)- space Turing Machine with high probability. We first observe that such simulation in selective protocols is stable. This means that selective protocols can be used to design efficient solutions to problems in O(log n)-space, which are also stable. In particular, we provide fixed-state stable efficient solutions to two central problems of leader election and majority computation, with confirmation. This is a separation result as stable efficient majority computation requires Ω( log n) states in standard population protocols, even if the leader is given. In addition, we study computation of the median in a comparison model where the operational state space of agents is fixed and the transition function decides on the order between (arbitrarily large) hidden keys associated with the interacting agents. We show that computation of the median of n numbers requires Ω(n) parallel time and the problem can be solved in O(n log n) parallel time in expectation and whp in standard population protocols. This is joint work with Adam Gańczorz, Tomasz Jurdziński and Grzegorz Stachowiak, Wrocław University

Venue

Bedford 0-07 [MS Teams Link]

Short Bio

Abstract

We study the (parameterized) computational complexity of problems in the context of fair allocations of indivisible goods. More specifically, we show fixed-parameter tractability results for a broad set of problems concerned with envy-free, Pareto-efficient allocations of items (with agent-specific utility functions) to agents. In principle, this implies efficient exact algorithms for these in general computationally intractable problems whenever we face instances with few agents and low maximum (absolute) utility values. This holds true also in high-multiplicity settings where we may have high numbers of identical items. On the technical side, our approach provides algorithmic meta-theorems covering a rich set of fair allocation problems in the additive preferences model. To achieve this, our main technical contribution is to make an elaborate use of tools from integer linear programming. More specifically, we exploit results originally going back to a famous theorem of Lenstra [Math. Oper. Res. 1983] concerning (the fixed-parameter tractability of) Integer Linear Programs (ILPs) with bounded dimension (that is, the dimension shall be considered as a (small) parameter) and the more recent framework of (combinatorial) N-fold ILPs. We reveal and exploit a fruitful interaction between these two cornerstones in the theory of integer linear programming, which may be of independent interest in applications going beyond fair allocations. We will also discuss some improvement using the Parametric ILP. Joint work with Robert Bredereck, Andrzej Kaczmarczyk, and Rolf Niedermeier

Venue

Bedford 0-07 [MS Teams Link]

Short Bio

Nicola is a Senior Lecturer in the Department of Informatics at King's College London. His interests are in safety and security assurance of cyber-physical (aka autonomous) systems, or CPSs, with an emphasis on biomedical applications. His research aims to develop formal analysis methods (verification, control, and synthesis) to design CPSs that are provably correct. With CPSs increasingly incorporating machine-learning components for e.g., sensing, control and model predictions, his work also focuses on data-driven verification of CPSs, whereby formal analysis and principled learning methods come together to provide correctness guarantees and interpretability, while accounting for the uncertainty and (potential) brittleness introduced by the learning components.

Abstract

We introduce PCFTL (Probabilistic CounterFactual Temporal Logic), a new probabilistic temporal logic for the verification of Markov Decision Processes (MDP). PCFTL is the first to include operators for causal reasoning, allowing us to express interventional and counterfactual queries. Given a path formula ϕ, an interventional property is concerned with the satisfaction probability of ϕ if we apply a particular change I to the MDP (e.g., switching to a different policy); a counterfactual allows us to compute, given an observed MDP path τ, what the outcome of ϕ would have been had we applied I in the past. For its ability to reason about what-if scenarios involving different configurations of the MDP, our approach represents a departure from existing probabilistic temporal logics that can only reason about a fixed system configuration. From a syntactic viewpoint, we introduce a generalized counterfactual operator that subsumes both interventional and counterfactual probabilities as well as the traditional probabilistic operator found in e.g., PCTL. From a semantics viewpoint, our logic is interpreted over a structural causal model translation of the MDP, which gives us a representation amenable to counterfactual reasoning. We evaluate PCFTL in the context of safe reinforcement learning using a benchmark of grid-world models.