Lightning Talk Abstracts - Confirmed

Jack Atkinson FTorch: A library for coupling PyTorch machine learning models into large Fortran codes (Institute of Computing for Climate Science)

Abstract: Many large modelling codes are seeking to incorporate recent developments in machine learning (ML). This is often done through ML working in conjunction with the existing numerical codes rather than as a complete replacement. Doing so introduces a number of challenges around reproducibility, re-usability, and language interoperation - ML development is usually conducted in Python whilst numerical scientific codes are often written in compiled languages like Fortran, C, or C++. These challenges form a barrier to researchers increasing complexity and ‘time-to-science’. We present FTorch, a library developed in the Institute of Computing for Climate Science (ICCS) at the University of Cambridge. The aim of this is to make it easy for researchers to develop ML models in the PyTorch framework and couple them to large numerical codebases in Fortran. Particularly in mind are those without an extensive computing background or access to software/computing support. Whilst its origins lie in the domain of weather and climate research we have since had interest from users in molecular modelling, astrophysics, civil engineering, and other scientific domains. https://github.com/Cambridge-ICCS/FTorch

Lloyd Fung Fast Bayesian Model Identification for Control (Applied Maths and Theoretical Physics)

Abstract: We pose the following problem in control theory: Assuming a control device has access to and control over the full state space of a dynamical system but does not have any prior knowledge of the governing equation of a dynamical system, what is the quickest and most optimal way for the control system to explore the state space and learn about the equation governing the dynamical system while controlling its desired state?
This is a generic problem for a lot of control engineering and may have real-world potential in applications such as drones and aircraft control when abrupt changes in the system require the controlling device to learn about new dynamics in real-time. It is also effectively a reinforcement learning problem, as the control device is required to learn about the system while balancing exploration and exploitation. However, different from many trendy reinforcement learning techniques, we constrain ourselves to models that can be written as dyanmical equations. To this end, the Sparse Identification of Nonlinear Dynamics (SINDy) framework has been shown to be effective in learning interpretable and parsimonious equations directly from data. However, it is not the most robust against noise and requires a lot of data to learn models properly. Our current work recasts SINDy in a Bayesian framework. By taking into account the uncertainty around the data, our method shows a more robust capability for learning the correct model in the low-data limit. Moreover, our method approximates the Bayesian likelihood and evidence, avoiding the need for computationally expensive Markov chain Monte Carlo (MCMC) sampling and enabling its use in real-time control applications.
By computing the potential information gain, the Bayesian approach can also inform optimal sampling and effectively form an active learning problem. The question remains, however, on how to incorporate this information into a control rule in optimal control theory. 

Mohammadreza Noormandipour  Solving Drug Discovery Optimization Problems with Ising Machines (Applied Maths and Theoretical Physics)

Abstract: In the rapidly evolving field of drug discovery, the ability to efficiently solve complex optimization problems is crucial for identifying novel therapeutic compounds, predicting drug-target interactions, and optimizing lead candidates. This talk will explore the application of Ising machines—powerful combinatorial optimization solvers—to key challenges in drug discovery. By mapping problems such as protein-ligand docking, molecular conformation optimization, and drug target interaction prediction onto the Ising model, we can leverage the unique capabilities of these machines to accelerate the discovery process. Specific use cases will include optimizing molecular conformations, designing new drug-like molecules, and predicting off-target effects, demonstrating how the combinatorial nature of these tasks can be effectively addressed using Ising-based approaches. Attendees will gain insight into the practical implementation of Ising machines and the potential for these tools to revolutionize the optimization landscape in pharmaceutical research.

Shrankhla Pandey A new software tool for the visualization, annotation, and segmentation of biomedical images (Astronomy)

Abstract: Mental health recovery narratives can provide recovery-oriented benefits [1], including increased quality of life and the presence of meaning in life. Our objective is to develop an automated system that can categorize recovery narratives with respect to seven dimensions, including genre, tone, and trajectory. We will achieve this by utilizing both natural language processing (NLP) and acoustic speech features, employing machine learning techniques.
[1] Slade et al. Recorded Mental Health Recovery Narratives as a resource for People Affected by Mental Health Problems.

Guillaume Proffit Historical Census Linking: issues of scale (History)

Abstract: Individual-level historical census for the decades 1851 to 1921 have been made avaialable to population researchers, and already been transformative. However, there is immense value in linking individuals across these censuses to follow evolution of individual or household chaarcteristics as causes or consequences of other factors (transports, demographic shifts, changes in the stucture of the economy, etc). Following litterature on American censuses, I have implemented a solution based on the Fellegi Sunter statistical model, but am facing issues with the scale of the computing: the worst case scenario to link two 30m censuses is to compare 9*10^12 pairs of records. These comparisons can be reduced thanks to blocking, but this risks creating false negatives. Even using other solutions such as an ML classifier, the core scale problem remains the same.

Zakhar Shumaylov AI Models Collapse When Trained on Recursively Generated Data (Applied Maths and Theoretical Physics)

Abstract: Stable Diffusion revolutionised image creation from descriptive text. GPT-2, GPT-3(.5) and GPT-4 demonstrated astonishing performance across a variety of language tasks. ChatGPT  introduced such language models to the public. It is now clear that large language models (LLMs) are here to stay, and will drastically change the ecosystem of online text and images. In this talk we consider what may happen to GPT-{n} once LLMs contribute much of the text found online. We find that use of model-generated content in training causes irreversible defects in the resulting models, where tails of the original content distribution disappear. We refer to this effect as model collapse and show that it can occur in LLMs, as well as in Variational Autoencoders and Gaussian Mixture Models. We build theoretical intuition behind the phenomenon and portray its ubiquity amongst all learned generative models. We demonstrate that it must be taken seriously if we are to sustain the benefits of training from large-scale data scraped from the web. Indeed, the value of data collected about genuine human interactions with systems will be increasingly valuable in the presence of LLM-generated content in data crawled from the Internet.

Jordan Skittrall Regions of interest in sequences of numbers, with application to finding novel biology in viruses (Pathology)

Abstract: Many approaches to finding runs of high or low values in sequences of numbers use sliding windows, which impose length scales on the problem.  These length scales are often arbitrary and in many problems it would be better to take an approach that makes no assumption about the length of a run of values of interest.  I was involved in a collaboration that developed a scale-agnostic approach to the problem of finding runs of extremal values, and have subsequently refined ideas in the approach to increase its discriminative ability.  I shall briefly describe the approach and how we have applied it to biological sequence analysis, predicting RNA structures important to the lifecycles of HIV, SARS-CoV-2 and influenza.  I am interested in discussions with anybody who might be interested in adapting the approach to their work.

Maha Waseem Leveraging AI for Personalized Early-Years Education: Prototypes and Potential (Judge Business School)

Abstract: In the rapidly evolving landscape of education, the integration of Artificial Intelligence (AI) holds immense promise for enhancing early-years teaching and caregiver support. This talk will delve into the innovative prototypes developed by Nesta, aimed at providing personalized and creative solutions for educators and caregivers of young children. We'll explore four key prototypes:Explain-Like-I’m-3: A web app designed to help educators and caregivers generate age-appropriate explanations for complex concepts, facilitating contingent talk with young children. Personalised Activity Generator: This tool offers tailored activity plans based on a child’s interests and age, drawing from trusted early-years learning guidelines. Activity Recommendation Engine: By interpreting early-years assessment forms, this engine suggests relevant activities from established resources like the BBC Tiny Happy People website, enhancing the practical application of assessments. Early-years Chatbot: A sophisticated AI-driven assistant that provides personalized education and parenting advice, addressing a wide range of caregiver needs while managing the challenges of knowledge base selection and misinformation. Join us to discover how these prototypes, rooted in interviews with early-years practitioners, aim to transform educational practices and support structures, fostering a more personalized and effective learning environment for young children. We will discuss the development process, the challenges encountered, and the potential impact of these AI tools on the future of early-years education.

Tristan Whitmarsh A new software tool for the visualization, annotation, and segmentation of biomedical images (Astronomy)

Abstract: ScanXm is a newly released software tool developed by Tristan Whitmarsh at the University of Cambridge. It aims to provide a simple and user-friendly interface for the annotation of 2D and 3D medical and biomedical images. ScanXm also includes several deep learning modules for image processing and the automatic segmentation of various organs and tissue types. Notably, this software requires no Python coding, will run without an expensive GPU, and does not use cloud computing, thus keeping your confidential patient data safe. Through a recent collaboration with NVIDIA, ScanXm now seamlessly integrates with MONAI Label. This integration enables ScanXm to run powerful AI models such as vision transformers. ScanXm now also links with Segment Anything 2 by Meta, which allows it to segment any object with only a few mouse clicks.

Organising Committee

Sireesha Chamarthi (Astronomy), Xin Du (Physics), Ed Harding (Clinical Biochemistry), Mohammadreza Noormandipour (Physics), Parley Ruogu Yang (Applied Maths and Theoretical Physics), Jun Jie Peng (Cancer Research UK), Anna Maria Papameletiou (Cancer Research UK), Maha Waseem (Judge Business School), Qiyao Wei (Applied Maths and Theoretical Physics), C.Zhang (Obstetrics and Gynaecology).

Registrations 

To attend, please register here (registration closes midday Weds 18th Sept)