MBZUAI and Berkeley explore the future of machine learning

Researchers from MBZUAI and University of California, Berkeley gave a glimpse of the future of machine learning, as they explored groundbreaking research and cutting-edge techniques that are shaping the fast-evolving field. 

The MBZUAI-Berkeley Joint Workshop on Emerging Directions in Machine Learning took place at MBZUAI’s campus in March, bringing together leading experts from both institutions to discuss their pioneering work. 

“Our main aim in organizing this workshop was to expose MBZUAI students and faculty to cutting-edge ideas, promote collaborative discussions, and help foster an environment of innovation,” said Samuel Horvath,  Assistant Professor of Machine Learning at MBZUAI, and one of the workshop’s organizers. 

“Workshops like this are crucial as they facilitate direct interaction between leading global experts and our students, enabling knowledge exchange, inspiring novel research directions, and encouraging collaboration.” 

The day saw eight talks and a panel discussion, with a speaker line-up that featured one of the most influential figures in the field, Professor Michael I. Jordan. 

“I’m happy to be here again at MBZUAI – this is maybe the sixth or seventh time I’ve come, and the growth rate is astounding,” said Jordan, who is Laureate Professor and Honorary Program Director at MBZUAI, and Pehong Chen Distinguished Professor in the Department of Electrical Engineering and Computer Science and the Department of Statistics at the University of California, Berkeley. 

“The topic of my talk was online learning, which is a methodology where you learn from arbitrary sequences and don’t make assumptions before you start the learning. It also applies to black box models – you can use it as a wrapper around a black box model.  

“That’s very much the spirit of current research – we have these very large, very complex models, and trying to mess around with the inside of them is basically impossible, whereas taking their outputs and giving them properties that are desired – for example being calibrated, being fair, being robust, and so on – that’s where research is going, and that’s what I talked about today.” 

Spearheading development 

During the morning session, Alireza Fallah (Berkeley) followed Professor Jordan’s presentation with a look his research ‘Fair Allocation in Dynamic Mechanism Design’, exploring fairness constraints in auction-based systems. Ian Waudby-Smith (Berkeley) then discussed ‘Anytime-valid Off-Policy Inference for Contextual Bandits’, introducing new statistical methods for analyzing adaptive learning policies.  

Before the midway break, Mingming Gong (MBZUAI) examined the theoretical aspects of causal inference in ‘On the Identifiability of ODEs/SDEs for Causal Inference’, focusing on how these models can be used to infer causal relationships from observational data. 

In the afternoon session, Maxim Panov (MBZUAI) introduced new techniques for improving prediction reliability in ‘Improving Conditional Coverage of Conformal Prediction Methods’, and Nils Lukas (MBZUAI) addressed security challenges in large language models with his talk ‘Leveraging Optimization for Adaptive Attacks against Content’, demonstrating how attackers can circumvent watermarking techniques.  

Martin Takáč (MBZUAI) highlighted the role of graph neural networks in accelerating scientific discovery in ‘Graph Neural Networks for Materials Discovery’, showcasing their applications in catalysis research. And Salem Lahlou (MBZUAI) provided an overview of generative flow networks in ‘GFlowNets: An Introduction and Recent Advances’, discussing their potential for structured exploration in AI-driven discovery. 

Reflecting on his talk about using AI to accelerate the discovery of new materials, Takáč, Deputy Department Chair and Associate Professor of Machine Learning, said: “I hope the students have been inspired to explore non-traditional applications of AI and machine learning. For me, that means discovering new materials or novel chemical pathways — but for them, it could mean innovation in entirely different ways.”

“When you bring experts from different areas, like we’ve done today with Berkeley, it can help students to broaden their view on AI. Many times, students work with their advisors and are very narrowly focused, so they are not challenged to look outside their field of interest. But there is great benefit to workshops like this to show them very different topics in one place. 

Alignment of purpose 

The workshop also served to strengthen the relationship between MBZUAI and Berkeley, which is ranked ninth in the Times Higher Education World University Rankings. 

“Partnering with Berkeley, one of the world’s premier universities renowned for pioneering research, underscores MBZUAI’s commitment to excellence and innovation,” said Horvath. “Collaborations of this nature significantly enhance our research capabilities, global visibility, and academic reputation. It also emphasizes our position among leading institutions worldwide, reflecting the calibre of our research and our shared dedication to advancing the frontiers of artificial intelligence and machine learning.” 

“There is a mission statement here at MBZUAI that I really resonate with,” added Jordan.  

“AI is viewed as an engineering discipline that’s built by real people and aids human thriving, aids the growth of employment, and aids the solutions to problems of this century including sustainability and social issues.  

“That’s also the mission of my group at Berkeley and that of my colleagues, so there’s a real alignment in the perspective on the development of this technology. It’s really great to share ideas and help to push that perspective further and faster.”