Choose timezone
Your profile timezone:
AI Collaborative Workshop
→
UTC
Maths Seminar Room (MB-503) (QMUL Mile End Campus)
Maths Seminar Room (MB-503)
QMUL Mile End Campus
Kostas Papafitsoros
(QMUL),
Marcella Bona
(QMUL),
Natasha Blitvic
(QMUL),
Peter Thorpe
(Queen Mary University of London),
Thomas Roelleke
(Queen Mary University of London)
Description
Contact
Participants
Amaranta Membrillo Solis
Amin Paykani
Axel Rossberg
Christos Vergis
Devis Di Tommaso
Ed Macaulay
Eleni Matechou
Enrico Camporeale
Evangelia Kyrimi
Iran Roman
Kostas Papafitsoros
Madeleine Ostwald
Martina Rama
Natasha Blitvic
Neil Cagney
Nikos Bempedelis
Patrick Healey
Rachel Parkinson
S.M.Hadi Sadati
Sara Mahdi
Søren Riis
Thomas Malcomson
Thomas Roelleke
Valentina Donzella
Vito Mennella
- +19
-
-
10:00
→
10:15
Introduction 15m Maths Seminar Room (MB-503)
Maths Seminar Room (MB-503)
QMUL Mile End Campus
Speakers: Kostas Papafitsoros (QMUL), Marcella Bona (QMUL), Natasha Blitvic, Thomas Roelleke (Queen Mary University of London) -
10:15
→
11:15
Session 1: Showcasing AI expertise 1h Maths Seminar Room (MB-503)
Maths Seminar Room (MB-503)
QMUL Mile End Campus
Talks by AI experts (from all schools; the term “expert” is understood in the broad sense) showcasing their expertise and potential applications to inspire collaborations.
-
“Humanity’s Last Exam”: What Hard Questions Teach Us About AI 15m
Humanity’s Last Exam is a cross-disciplinary benchmark built from expert-written problems in computer science, mathematics, and beyond. As a contributor, I’ll share insights from designing questions that remain challenging for state-of-the-art AI models—without revealing the protected items. Through analogous examples and recurring failure patterns, I’ll show where today’s reasoning systems excel and where they break. I’ll conclude with practical ideas for collaboration across disciplines.
Speaker: Søren Riis (QMUL) -
Finding shape and structure in data: geometry and topology in the natural sciences 15m
Geometric and topological tools in data analysis are changing the way we look at data in the natural sciences. In this presentation, I’ll show how these methods can reveal patterns and shapes hidden in complex datasets. We’ll see examples from chemistry, soft matter, and nanomaterials, where geometric and topological tools help us describe structure and behaviour beyond what standard data analysis techniques can capture. I’ll also point to some new and promising applications, showing how geometry and topology can give a fresh perspective on problems across the natural sciences.
Speaker: Amaranta Membrillo Solis (QMUL) -
Applying machine learning for catalysis discovery 15mSpeaker: Devis Di Tommaso (Queen MAry University of London)
Electrochemical CO2 reduction (CO2R) is a promising carbon capture and utilization strategy for producing value-added chemicals such as ethylene and ethanol. A key challenge in CO2R research is designing catalysts that selectively convert CO2 into specific products. In this talk, I will present an overview of computational strategies we have developed, combining high-throughput calculations and machine learning to identify potential CO2R catalysts. These approaches include the use of machine learning potentials to model the structure of highly complex catalytic systems and graph neural networks for predicting catalytic properties.
-
Do we need better AI or better data? 15m
Intelligent Systems rely on the relationship between data collected by different types of sensors and the machine learning algorithms using the data to build understanding and knowledge, i.e. ‘situational awareness’. Very often, more accurate ‘situational awareness’ is associated with bigger and deeper neural networks, or more data used in training. However, this paradigm can be changed: can we better understand the relationship between data, its unique features, its degradation, and AI?
Speaker: Valentina Donzella (QMUL)
-
-
11:15
→
11:45
Group Discussion and Coffee (Maths Common Room) 30m Maths Common Room
Maths Common Room
QMUL Mile End Campus
-
11:45
→
13:00
Session 2: Projects potentially benefitted by AI 1h 15m Maths Seminar Room (MB-503)
Maths Seminar Room (MB-503)
QMUL Mile End Campus
Researchers from various disciplines outlining their work and where AI expertise could advance it.
-
Artificial Companions and Relational Intelligence. 15m
Digital companions are emerging as one of the fastest growing and most controversial application areas for generative AI. This talk will highlight their potential for transforming research in the science of social interaction.
Speaker: Patrick Healey (EECS, QMUL) -
Imbalanced regression with applications in space physics 15m
Classical machine learning methods often struggle with strongly imbalanced datasets, particularly in regression problems involving rare and extreme events where the distribution of the target variable is highly skewed. While numerous techniques exist for classification, imbalanced regression remains relatively under-explored. Contrary to the prevailing deep learning paradigm that “more data is better,” we argue that, in such cases, the most informative data (those capturing extremes) are more valuable than abundant, redundant samples.
In this context, I will present three complementary methods with applications to space weather prediction. ACCRUE (Accurate and Reliable Uncertainty Estimate) is a model-agnostic, post-hoc approach that converts deterministic models into probabilistic ones by ensuring that their uncertainty estimates are both accurate and reliable. PARIS (Pruning via Analysis of Representer-based Influence Scores) provides a principled framework for influence-aware dataset reduction, using the representer theorem to identify and remove redundant or detrimental samples, thereby improving model performance on extreme events. Finally, ProBoost (Probabilistic Boosting) integrates ACCRUE and PARIS in an uncertainty-weighted ensemble, where each model contributes proportionally to its calibrated confidence, offering a robust, probability-based alternative to conventional boosting.
Although demonstrated in the context of space weather, these methods are broadly applicable to any domain focused on modeling rare or extreme phenomena.Speaker: Enrico Camporeale (QMUL) -
Nonequilibrium many-body physics with statistical ensembles 15m
Statistical mechanics uses ensembles to describe and predict equilibrium properties of macroscopic systems. An ensemble collects microscopic configurations of the degrees of freedom that share a few macroscopic characteristics (e.g., energy, volume, number of particles), and the fundamental postulate is that other equilibrium properties will also be shared by the ensemble members. This idea is based on the phenomenological fact that we can describe and control macroscopic properties of large systems without knowing all their microscopic details. But macroscopic regularity and stability are phenomenologically observed away from equilibrium as well, e.g., in relaxation processes or driven systems. Yet for nonequilibrium processes, no similar theoretical connection from microscopic laws to macroscopic behavior is known. Broadly speaking, the main challenge is that there is a lot more going on away from equilibrium as macroscopic observables change over time, concealing fundamental patterns and complicating the attribution of effects to causes. Given their demonstrated and constantly improving capabilities to uncover patterns in large, complex datasets and to compress information by extracting relevant latent variables, for example, machine learning methods appear well suited to overcome this challenge.
Speaker: Lennart Dabelow (Queen Mary University of London) -
3D segmentation of airway tissue and cells content using AI models 15m
Nanoscale imaging technologies, such as Volume Electron Microscopy (Volume EM), are revolutionising our understanding of tissue architecture in situ. These approaches have already enabled large-scale analyses of synaptic connections and the identification of novel cellular contacts and subcellular structures.
Furthermore, nanoscale information can now be integrated with microscale data obtained from complementary imaging modalities such as tissue micro-CT and X-ray imaging, allowing the development of a multiscale understanding of tissue organisation in both healthy and diseased states.
Biological discovery increasingly depends on accurate image recognition and 3D segmentation of complex biological structures. However, these processes still rely heavily on manual segmentation or refinement of datasets produced by AI-based algorithms.
In this presentation, we will discuss the opportunities and challenges associated with applying AI algorithms to 3D segmentation analysis in electron microscopy, highlighting current advances and future directions for automated biological image analysis.Speaker: Vito Mennella (Queen Mary University) -
Towards Intelligent Endoluminal Navigation & Intervention: Theoretical, Artificial, and Embodied Intelligence in Action 15m
At ACEi group (Artificial, Collective, Embodied intelligence), we envision revolutionizing endoluminal catheterization through the integration of soft robotics, embodied intelligence, and artificial intelligence. This innovation aims to enable minimally invasive interventions deep within the body via automated multi-modal tissue and endoluminal navigation.
ACEi’s vision is to make multi-modal navigation for deep body access possible. Robotic Mechanical Thrombectomy and Neuroendoluminal catheterization are key clinical applications that underpin this vision. This talk presents an overview of our theoretical, artificial, and embodied intelligence frameworks, demonstrating their application in intelligent endoluminal navigation and intervention.Speaker: Hadi Sadati (SEMS, QMUL)
-
-
13:00
→
14:00
Lunch 1h Maths Common Room
Maths Common Room
-
14:00
→
15:00
Session 3: More ideas on AI/ML applications and proposals for collaborations and future developments 1h Maths Seminar Room (MB-503)
Maths Seminar Room (MB-503)
QMUL Mile End Campus
A session continuing the presentation of AI/ML applications, keeping an eye on possible collaborations and future developments. We welcome discussions on previously unsuccessful grant bids, targeting their improvement by adding elements of AI towards a potential resubmission, possibly on short time-scales.
-
Multimodal AI for Personalized Health Management 15m
The integration of artificial intelligence (AI) into personalized holds the potential to revolutionize personalized health management. With the advent of modern technology, it is now possible to acquire real-time signals via wearables and wireless sensors which enable individuals to track their well-being, allowing them to adjust their daily routine based on their underline comorbidities and chronic conditions (e.g. diabetes, cancer etc.). Research is being conducted based on design and development of novel multimodal techniques which allow training the real-time signal variations based on underline comorbidities acquired through medical testing, electronic health records, imaging techniques etc. However, despite technological advancements, the design and development of multimodal AI encounters challenges such as information fusion, prediction explainability, practicalities in real-time implementation etc.
In this talk, I will discuss these challenges, cutting-edge research, innovative applications, and successful case studies of multimodal AI design. Participants will gain a comprehensive understanding of the role of multimodal AI in personalized health management and precision medicine.Speakers: Muhammad Salman Haleem (Queen Mary University of London), TBA -
Can Visual Language Models benefit Wildlife Re-Identification? 15m
Wildlife Re-Identification (re-ID) denotes the task of identifying individual animals from images, based on their morphological patterns. This tool forms the backbone of many wildlife conservation projects, and informs research spanning from behavioural studies to extinction risk assessments. Current AI-based models for animal re-ID are based on deep metric learning and operating under an image retrieval setting. Here we would like to explore whether large Visual Language Models (VLMs) can also be applied to the task. Preliminary experiments have revealed that state-of-the-art VLMs...completely fail on animal re-ID! How can we make them work?
Speaker: Kostas Papafitsoros (Queen Mary University of London) -
Extracting self-similarity from data 15m
Identifying self-similarity is key to understanding and modelling a plethora of phenomena across a range of disciplines. Unfortunately, this is not always possible to perform formally in highly complex problems. We propose a methodology to extract the similarity variables of a self-similar physical process directly from data, without prior knowledge of the governing equations or boundary conditions, based on an optimisation problem and symbolic regression. We analyse the accuracy and robustness of our method in problems which have been influential in fluid mechanics research; the algorithm recovers the known self-similarity expressions in the first four problems and generates new insights into single length scale theories of homogeneous turbulence.
Speaker: Nikos Bempedelis (Queen Mary University of London)
-
-
15:00
→
15:30
Group Discussion and Coffee (Maths Common Room) 30m Maths Common Room
Maths Common Room
-
15:30
→
16:00
Teams & Collaboration Workshop (Maths Common Room) 30m Maths Common Room
Maths Common Room
-
16:00
→
16:15
Plans going forward 15m Maths Seminar Room (MB-503)
Maths Seminar Room (MB-503)
QMUL Mile End Campus
-
10:00
→
10:15
