Cross-Hub collaboration takes shape after methodology workshop
Researchers exhibited posters on a range of topics relating to Generative AI methodology at the Hub X event
Hub researchers dodged tube strikes and rain to join colleagues from the Prob_AI Hub and ELLIS for a two-day workshop aimed at deepening cross-hub ties on core methodology.
The event, organised by Hub members and Methodology co-leads Arthur Gretton (UCL Gatsby) and Siddharth N. (Edinburgh), brought together researchers to map out shared ground and identify where the hubs could make the biggest joint impact. Three areas emerged as the strongest candidates for future collaboration: Gen AI for Science, one-step and few-step generative AI, and improved diffusion models.
A strong line-up of speakers set the scene with the latest methodological developments:
· Richard Turner of the University of Cambridge exposed the technical blind spots of major weather forecasting models, from high training costs to poor sensitivity to outliers.
· Arnaud Doucet, of Google DeepMind presented SURF, an unsupervised audio-cleaning method that sidesteps the scarcity of clean training data and matches supervised generative baselines.
· David Pfau, of Google DeepMind introduced his 2025 ICML paper on Wasserstein Policy Optimization, with applications in robotics, and drew out the surprising connections between quantum physics and reinforcement learning.
Richard Turner pulls back the curtain on AI Earth Systems Forecasting, highlighting the critical technical problems often swept under the carpet by major medium-range weather models
The workshop also shone a light on live collaboration opportunities within the Gen AI Hub. Hub director David Barber (UCL) issued a direct call for collaborators to help build a one-step generator from scratch - no diffusion model required. "We'd love to show the world that you don't need diffusion models... we want this to be the diffusion model killer," he said.
That project, co-led with Siddharth N., is one of several currently seeking collaborators. Browse the full list and find out what expertise each team needs on the Live Project Board.
Siddharth reflected on what the workshop achieved:
"The workshop was the first step towards building stronger ties between common strands in the UKRI GenAI and Pro-AI Hubs, discussing foundational challenges in methods and algorithms.
“The talks and discussions have strengthened ties and set the groundwork for further targeted collaboration towards a position paper and subsequent wider grant proposal."
Generative AI Projects in Progress: a spotlight on two of our research collaborations
Join us on 2 July at 11am (BST) for a free webinar spotlighting two Gen AI Hub collaborative projects. The lead researchers will share their work in progress and explain the benefits and challenges of managing a multi-partner collaborative project. There will be an opportunity for questions.
