The UKRI Inter AI CDT Conference took place on the 30th and 31st October 2023 at the Bristol Hotel with the Interactive Artificial Intelligence CDT based at the University of Bristol, the Foundational Artificial Intelligence CDT based at UCL and the CDT in ART-AI based at the University of Bath. Over 180 students, academics and industry representatives came together for this 2 day conference with the aim of enabling collaboration as well as delivering pertinent workshops.
On day 1 we had Jasmine Grimsley, from The London Data Company as the key note speaker, followed by a choice of 2 x morning and 2 x afternoon sessions (details of these sessions are below). At the end of day 1 we had a poster session at the MShed, which had over 40 posters submitted by students across the 3 x CDTs. Prizes for the best posters from each CDT were presented to the students at the conference dinner. Congratulations go to Dan Beechey, from ART-AI cohort 3, who won the competition from ART-AI!
On day 2 we had Jakob Zeitler, Founder of Matterhorn, as the key note speaker and this was followed by a morning session and 2 x afternoon sessions (details of these sessions are below). To close the conference we had a key note from Steven Schockaert from Cardiff University.
Below is some of the feedback we have received by those that attended:
“Apart from the networking, I really loved the opening talk from the London Data Company – it was a really good example of a company with very close ideals to ART-AI”
“I really enjoyed the poster session and the vision talks as they were the most relevant to my day to day research.”
“I enjoyed the climate event as I don’t feel that is talked about enough in our space or at previous events.”
“I really enjoyed getting to socialise with the extended CDT community and discuss my research at the poster session.”
“The poster session was amazing. it was the perfect opportunity to talk to others.”
If you would like to see more about what happened at the conference and to hear from some of the students, please watch the video below:
Information about the sessions
Day 1 – Morning
AI Vision – The William Jessop Suite
Prof Lourdes de Agapito (UCL) will be introducing 3 keynote speakers discussing their research in the field of AI Vision.
Christian Rupprecht (University of Oxford)
Title
Unsupervised Computer Vision in the Time of Large Models.
Abstract
With larger and larger models trained on billions of images (and sometimes text) entering the research landscape of computer vision is changing. The lines between unsupervised, few-shot and supervised learning are becoming blurry as using these larges models introduces information at a scale that is very difficult to assess and categorize. In this talk we will analyse the current state of the field, future research directions and some current practical applications with and without the use of large models.
Bio
After completing his BSc and MSc at the Technical University of Munich, Christian Rupprecht obtained his PhD in 2018 advised by Nassir Navab (Technical University of Munich) and Gregory D Hager (Johns Hopkins University). He joined Oxford University first as a PostDoc in Engineering Science with Andrea Vedaldi, then as a Departmental Lecturer in Computer Vision, and now an Associate Professor in Computer Science and Tutorial Fellow at Magdalen College.
Ed Johns (Imperial College London)
Title
Vision-Based Robot Learning of Everyday Tasks
Abstract
Most of the major recent breakthroughs in AI have emerged from training neural networks on huge datasets of images and/or text. However, currently we do not have the equivalent scale of data for robotics. To address this, my team and I have been developing very data-efficient methods for robots to learn new tasks, such as hammering in a nail, inserting a plug into a socket, and setting a table for dinner. In this talk, I will take you through a series of projects which have explored this, with two different strategies. The first strategy is through the use of human demonstrations, where we have been studying different modes of generalisation, such as generalisation to novel object poses, generalisation to novel object shapes, and generalisation to novel object grasps. The second strategy is through the use of pre-trained web-scale models, such as image diffusion models and vison-language models, which we are beginning to see can enable a range of tasks to be performed by a robot in a zero-shot manner, without requiring any demonstrations at all.
Bio
Dr Edward Johns is the Director of the Robot Learning Lab at Imperial College London, where he is also a Senior Lecturer. He received a BA and MEng from Cambridge University, and a PhD from Imperial College. Following his PhD, he was a post-doc at UCL, before returning to Imperial College as a founding member of the Dyson Robotics Lab, where he led the robot manipulation team. In 2017, he was awarded a Royal Academy of Engineering Research Fellowship, and then in 2018, he was appointed as a Lecturer and founded the Robot Learning Lab.
Laura Sevilla-Lara (University of Edinburgh)
Title
Efficient Video Understanding
Abstract
Video understanding is a fundamental ability for intelligent systems, from virtual assistants to robots or self-driving cars. Like in most learning problems, the trend in our research community has been to use larger models and larger datasets to improve our methods. While this has helped make progress, it is starting to catch up with us: it takes us months to train some models, and we are often restricted to short videos. In this talk I will present several principles that we have used in our lab to make video understanding technology less dependent on large amounts of data or compute, crucially without sacrificing accuracy.
Bio
Laura Sevilla is a lecturer at the University of Edinburgh since 2019, where she leads the video understanding lab. Before Edinburgh, she was at Facebook Research in California, where she worked in the video team. Before then, she was a postdoctoral researcher at the Max Planck Institute in Tuebingen, working on optical flow and action classification. She obtained her PhD in 2015 from the University of Massachusetts Amherst, working on object tracking, optical flow and applications to computational photography. She received the Google Faculty Research Award in 2020 and the Google Research Scholar Award in 2022. In addition to video understanding, her other body of work has been in outreach and broadening the scope of vision, through the organization of a series of workshops on computer vision for global challenges
Reproducibility – Ballroom
Best Practices in Research Software Engineering and ML
Presenters – Christopher Woods and Daladier Sampaio Neto (University of Bristol)
Sharing your scripts or code with others can be scary. It may run on your computer today, but how can you trust that it will work correctly somewhere else or at another time? We can raise similar questions about machine learning models or experiments. How to ensure that the experiments results are reproducible? How to track the best models from multiple experiments? How to keep track of a deployed model? We’ll share some of the best tactics in research software engineering and ML-Ops that will give you the confidence to share your software and model results with others. We will show you how you can trust that your code will give the same answers on other computers or at future points in time. We’ll reveal the best practices that will help you turn the code you write into software or models that can be shared and developed by your research community.
Day 1 – Afternoon
NLP talks – The William Jessop
Chaired by Dr Harish Tayyar Madabushi, University of Bath. Confirmed speakers:
- Elena Kochkina, JPMorgan AI Research
- Michael Schlichtkrul, University of Cambridge
Michael Schlichtkrul, University of Cambridge
Title
Taking Automated Fact-checking to the Real World
Abstract
Automated fact-checking is typically presented as an epistemic tool that fact-checkers, social media consumers, and other stakeholders can use to fight misinformation. Nevertheless, most research papers are surprisingly vague about how this technology is to be used. In the first part of this talk, I will argue that vague arguments about intended use hinders research in the area. I will use content analysis of highly cited papers to document and clarify the problem, and to establish recommendations. In the second part, I will try to follow my own recommendations, as I introduce a new dataset for automated fact-checking. I propose that human fact-checking is an effective method for fighting misinformation, and accordingly attempt to reverse-engineer the process. The resulting dataset allows reasoning about the capacity for models – including LLMs – to help human fact-checkers with some or all of their real-world fact-checking tasks.
Bio
Michael Schlichtkrull is an affiliated lecturer and postdoctoral research associate at the University of Cambridge, where he works on automated fact checking and other epistemically complicated NLP problems. Michael received his PhD from the University of Amsterdam, where he worked on graph neural networks for NLP, tackling problems including relational link prediction, question answering, and interpretability.
Elena Kochkina, JPMorgan AI Research
Title
Rumours and Opinions: Social Media Language Processing
Abstract
Social media data provides a valuable source of insights into human behaviour and societal trends. The possibilities for research on social media platforms are extensive, including sentiment analysis, hate speech detection, social network analysis, political research, and more. Elena will talk about her research on Tackling Online Misinformation in Social Media Conversations using Natural Language Processing, Social Media data collection and annotation, Temporal Persistence of Social Media classifiers.
Bio
Elena Kochkina is an AI Research Scientist at JPMorgan Chase &Co. Previously she was a Postdoctoral Researcher at Queen Mary University of London and the Alan Turing Institute. She holds a PhD from the University of Warwick. Her expertise is in Tackling Online Misinformation using Natural Language Processing. Her research interests include NLP Applications in Finance, Fact-checking and Model Generalisability from the temporal and domain perspectives.
Bristol student-led session – Ballroom
ChatGPT search in own data
An application and workflow that enables natural language search and synthesis own data through:
- preprocessing of local text repositories
- local filtering of relevant text through keyword
- search in subset of highest ranking documents with ChatGPT
The workshop will include a presentation of the main concepts, a demonstration, a practical hands-on session, and a discussion on risks and opportunities for deployment in organisations.
Day 2 – Morning
AI in energy and environment
Chaired by Professor Lorraine Whitmarsh, University of Bath. Three invited speakers will talk for 25 minutes each, with questions from the floor, about how they use AI to address environmental issues/questions in their work. This will be followed by a wider panel discussion on future directions and challenges for AI and the environment. Confirmed speakers:
- Chris Budd, University of Bath: Predicting the impact of climate change on the energy network
- Jatinder Mehimi, Environment Agency
- Travis Coan, University of Exeter: Big data analysis of social media
Day 2 – Afternoon
UCL student-led session
Making AI Safe
An interactive session brainstorming the risks of AI and strategies to mitigate them with Reuben Adams & Robert Kirk
Do you think AI is going to kill us all, or do you think there’s nothing to worry about? Somewhere in between? Come to this interactive session where you’ll try and understand in more detail what you think will happen. Working with others, you’ll brainstorm risks from AI, propose solutions and then try to find ways in which those solutions won’t work. Through this “builder-breaker” workshop you’ll build a more detailed understanding of the potential risks from AI and how optimistic you are about solutions to them.
Bath student-led session
BCI workshop with NeuroCONCISE
