Investigating collaborative learning at MBZUAI’s AI Quorum

Original article published in “Wired Middle East” on October 24, 2022

IT NEVER HURTS to join minds to create something phenomenal, the more, the merrier. This was the model of the historic Solvay Conferences which put some of the most brilliant minds to work on the most challenging physics and chemistry problems of the day. A century later, that’s exactly what the AI Quorum launched and hosted by Mohamed Bin Zayed University of Artificial Intelligence (MBZUAI) is working towards.

No surprise that this event takes place at MBZUAI as, geographically, Abu Dhabi serves as the gateway to three continents home to a majority of the world’s AI consumers. A winter series with a goal of stimulating AI research with leaders in the field of science and technology, the AI Quorum focuses on curiosity, collaboration and authenticity in assembling a research agenda that can be utilized to imagine the endless possibilities of AI. The series spans across October of 2022 to March of 2023, with one hybrid session every month.

The first of the sessions took off for October, with MBZUAI’s Professor Michael Jordan leading the meeting with around 20 noted research experts on the topic of collaborative learning in using and building technology for all to use. “Collaborative learning is basically what happens at the edge of a very large scale system where there’s learning happening. The edge devices might be cell phones, it might be hospitals, or any that have data. And so, they’re supplying the data to a central site who is using the data with machine learning to build a large model, that hopefully is better than the model that any individual could build,” explains Professor Jordan.

Jordan mentions that this type of learning also poses its own risks that all should be aware of and try to mitigate. One such issue is free-riding of data, whereby people may see that others are already supplying data and are not incentivized to provide any of their own. Similarly, there could also be adversaries trying to provide fake data or bad data to mess up the system. Separately, worrying about heterogeneity also comes up, “you don’t necessarily try to build one model for everybody. Maybe you want different models for different kind of subgroups of people,” offers Jordan.

Now, it is acknowledged that the up-and-coming technologies involving cloud and sky computing allow for this kind of supplying of data, but Jordan stresses that collaborative learning isn’t just about that. “It’s not just they’re supplying data, and then someone builds a nice language model. Maybe they’re trying to solve a local problem or making some weather predictions locally or resource allocation or design problems. And they have some data, but also nearby entities have data,” he shares.

From the successful event also comes a position paper in the works at MBZUAI. The paper will cover what the right design of an emerging system is so that “it’s not just about cloud computing or about sky computing or about centralized computing, it’s also about computing at the edge that reflects the fact that you have agents and people at the edge who have utilities, legal constraints and privacy issues,” Jordan mentions. 

“There is some treatment of edge devices, GPUs and CPUs. But really, there’s not been very much treatment of system support for edge devices of this kind wherein data flows and edge devices then supply models and predictions back to the edge,” says Professor Jordan, predicting that within the next five years to a decade there will be a turnaround for this as a major agenda item. 

This is only the beginning.