INFORMED-AI Hub

Applications are now open for entry in academic year 2026/27 for the Centre for Doctoral Training in Fundamentals of AI University of Oxford with Ellison Institute of Technology Oxford: https://lnkd.in/dakKvTxx

✳️ We are hiring ✳️ Postdoctoral Research Associate in Information Theory and Statistics (Fixed Term) The PDRA's main project will involve investigating several theoretical questions concerning distributed testing and estimation, particularly in streaming settings where agents have limited memory. They will study tradeoffs between memory, accuracy, and sample complexity of fundamental questions in hypothesis testing. In addition, the PDRA may be responsible for developing and conducting collaborative research projects as part of the overall work of the INFORMED-AI programme. The successful candidate will be based in Cambridge, primarily hosted by Varun Jog and co-supervised by Po-Ling Loh (Cambridge) and Sidharth Jaggi (Bristol). Full description and how to apply: https://lnkd.in/eFn7ffpF For further information or an informal chat about the INFORMED AI Hub, please get in touch [email protected]. We'd love to hear from you. Faculty of Mathematics, University of Cambridge EPSRC

📣 Applications Now Open for the Quantum Informatics CDT (QI CDT) 📣 The Quantum Informatics CDT is now inviting applications for up to 16 fully funded four-year PhD studentships beginning in September 2026. Quantum computing offers revolutionary ways to tackle the world’s most complex challenges. Realising its potential requires a new kind of expertise that bridges hardware, software, and applications. The QI CDT brings together five leading universities - The University of Edinburgh, Heriot-Watt University, UCL, the University of Oxford, and the University of Strathclyde - to train the next generation of researchers who will help define the future of this field. This unique four-year programme combines advanced research with integrated study across computer science, mathematics, physics, and engineering. Students benefit from a taught first year that builds foundational skills, followed by an in-depth research project. Supported by over 70 academic supervisors and more than 30 industry partners, the programme includes opportunities for placements, leadership and entrepreneurship development, and collaboration across academia and industry. Funding covers tuition fees, a four-year stipend, and a personal budget for research and travel. Both UK and international candidates with strong academic backgrounds in relevant disciplines are encouraged to apply. The application deadline is 8th January 2026 🎓 Find out more and apply via the QI CDT website: https://lnkd.in/eZ8mZhus Be part of the community shaping the future of quantum technologies. #QuantumComputing #QuantumInformatics #PhDOpportunity #PhD #Research #CDT #STEMCareers #PhDStudentships #QuantumTechnology Chris Heunen Ben Leimkuhler Erika Andersson Aleks Kissinger Mark Parsons Luigi Del Debbio Majid Safari Elham Kashefi

📢 New research alert: The NHS elective waiting list in England must be halved to meet the 18-week treatment target. A new modelling study by Hub academic Neil Walton and colleagues Richard M. Wood, Sebastian S. Fox, Lucy E. Morgan, (published in Future Healthcare Journal) finds that the current list of 7.5 million patients must fall to 3.4 million to restore the NHS constitutional standards. 📈 The work undertaken by these researchers is a vital contribution to understanding the scale of the elective care backlog and the policy, workforce, and operational changes needed to address it. 🔗 Read the full article here: https://lnkd.in/dmFrz5Ek EPSRC NHS England Durham University Business School #HealthcarePolicy #ElectiveCare #WaitingTimes #HealthInequality #FutureHealthcare #HealthModelling

Research Update: New results on the Entropy Power Inequality Hub researcher Lampros Gavalakis with colleagues Matthieu Fradelizi (Uni Gustave Eiffel) and Martin Rapaport (Carnegie Mellon University) presented their research on A dimensional improvement of the Entropy Power Inequality under marginal assumptions at the 2025 IEEE International Symposium on Information Theory ISIT 2025 in Michigan this summer.  https://lnkd.in/dyryQEVF Their paper presented classic results from geometry reimagined through the lens of information theory. 📊 They strengthen the well-known Entropy Power Inequality (EPI) and Fisher Information convolution Inequality in higher dimensions. 💡 Why it matters: The entropy power inequality (EPI) is a cornerstone of information theory, with far-reaching applications in data compression, signal processing, and machine learning. Marginal assumptions are often realistic in practice, and when the sample size is large, dimensional improvements can make a significant difference. Full article🔗https://lnkd.in/exRaVxkY Lampros and Ioannis Kontoyiannis (University of Cambridge) have also been working on “Conditions for equality and stability in Shannon’s and Tao’s entropy power inequalities”   https://lnkd.in/exzQP-QR The paper proves that equality in Shannon’s EPI holds if and only if the variables are Gaussian under the weakest possible assumptions, justifying de Bruijn’s identity without second-moment requirements, and it also establishes new stability results (including for Tao’s discrete EPI). The EPI has had immense impact in communications since the beginning of the field and de Bruijn’s identity has found a number of applications including MMSE of estimators. This work shows that it is mathematically justified and identifies the weakest possible assumptions. In addition, applications to generative models differentiate the entropy of diffusions starting from discrete distributions; this work shows that this is always justified for Gaussian perturbation. University of Cambridge EPSRC 

INFORMED AI Seminar Series Autumn 2025/26 We are delighted to announce the INFORMED AI Hub’s seminar series for Term 1 of 2025/26 academic year. The schedule of talks is detailed online: https://lnkd.in/eSXtrQyW The next one is Tuesday 7 October 2025: Yingzhen Li - Variational Uncertainty Decomposition for In-Context Learning https://lnkd.in/eVTPHwA3 Events are open to all interested parties and will be hybrid to enable in person or online attendance. The seminar/ talk will run from 2.00 – 3.00pm followed by 30 minutes for questions.   Please get in touch with the organisers via  [email protected] to express your interest in joining online or in person for any or all of these seminars. Jianhong Wang Imperial College London EPSRC University of Bristol Faculty of Science and Engineering Please keep checking back for more abstracts and titles as they are confirmed.

📢 EDINBURGH PUBLIC FORUM - AI IN THE CITY 📆 30th October 🕰️ 6pm-7.30pm 🏰 Edinburgh Futures Institute AI For Collective Intelligence (AI4CI), Architecture and Design Scotland (A&DS) and Edinburgh Futures Institute (EFI) are pumped to be hosting a #FREE #PublicForum on ‘AI in the City’ in historic #Edinburgh. Join us for a broad and rounded discussion between expert panellists Prof. Alison Heppenstall, Caroline Parkinson, Jeremy Doherty, Raffaele Esposito and Tom Wilkinson on how they see AI affecting our towns and cities. We are delighted to announce that alongside our great panel this event will also be professionally moderated by Scottish journalist, broadcaster and presenter #StephenJardine. This event promises to give members of the public access to expert knowledge and the opportunity to ask the questions they most want the answers to. Join us on 30th October and stay engaged with the influence of AI on city living, building and planning; the challenges, the opportunities, the risks, the rewards and the impact to communities living in these areas. Get your FREE tickets here 👉 https://lnkd.in/eMn7XZkN or scan the QR code. You can learn more about the forum and our panellists by visiting our website 👉 https://lnkd.in/eGDtk9yB #PublicForum #AI #SmartCityDesign #Scotland #FreeEvent #AIintheCity #FuturePlanning #Edinburgh

Many congratulations to all the authors here. We're thrilled that INFORMED AI is represented here with the work of Jianhong Wang and his colleagues. University of Bristol Faculty of Science and Engineering EPSRC

We're looking forward to hearing Prof Moshe Vardi talking on 'Efficiency, Resilience, and Artificial Intelligence' when he delivers the 2025 Wheeler Lecture on Wednesday 22 October. Moshe Y Vardi is University Professor and the George Distinguished Service Professor in Computational Engineering at Rice University. His lecture starts at 15:00 (BST) and will also be streamed live for those unable to attend in person. Interested in attending? Then please register here: https://lnkd.in/eziNGQwg

Excited to share our new preprint: “Bilateral Distribution Compression: Reducing Both Data Size and Dimensionality” Training AI models often now requires enormous datasets, computation, and energy, with serious financial and environmental costs. A key challenge is how to reduce data without losing the essential information it carries. With my supervisors Nick Whiteley, Robert Allison, and Tom Lovett, we introduce Bilateral Distribution Compression (BDC), a framework that compresses datasets in both sample size and dimensionality (i.e. bilaterally) while preserving their underlying distribution. - Traditional distribution compression (e.g. Kernel Herding) shrinks datasets by sample size only. - Dimensionality reduction (e.g. PCA, autoencoders) shrinks features but leaves sample counts unchanged. - Our method does both, with linear scaling in dataset size and dimension. A key idea is to break a hard problem into two easier steps, using three versions of the Maximum Mean Discrepancy (MMD): - Reconstruction MMD (RMMD): checks if an autoencoder’s reconstructions preserve the distribution of the data (not just pointwise distances). - Encoded MMD (EMMD): checks if the compressed latent set represents the encoded data well in latent space. - Decoded MMD (DMMD): the final goal, checks if the decoded compressed set still looks like the original dataset. The intuition: if the autoencoder preserves the data distribution (low RMMD) and the latent compressed set matches the encoded data (low EMMD), then the decoded compressed set must also match the original distribution (low DMMD). In practice, BDC achieves very significant reductions, surpassing 99.99% compression on some tasks, while matching or outperforming existing compression approaches across various regression, classification and clustering problems, all at a fraction of the cost. Preprint: https://lnkd.in/e4EEdKqj A big thank you to my supervisors for helping bring this idea to life in such a short time frame!