The early years of faculty members’ careers are a formative and invigorating time in which a solid foundation can be established that will aid determine the trajectory of researchers’ studies. This includes building a research team that requires inventive ideas and direction, inventive collaborators, and reliable resources.
For a group of MIT faculty working in and around artificial intelligence, early collaborations with the MIT-IBM Watson AI Lab on projects have played an critical role in helping to promote ambitious research directions and shape prolific research groups.
Building dynamics
“The MIT-IBM Watson AI Lab has played an extremely important role in my success, especially when I first started out,” says Jacob Andreas—an associate professor in the Department of Electrical Engineering and Computer Science (EECS), a member of MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL), and a researcher at the MIT-IBM Watson AI Lab—who studies natural language processing (NLP). Shortly after joining MIT, Andreas began his first major project at the MIT-IBM Watson AI Lab, working on language representation and structured data augmentation methods for low-resource languages. “It really allowed me to get the lab up and running and start recruiting students.”
Andreas notes that this occurred at a “pivotal moment” when the field of NLP was undergoing significant shifts toward understanding language models – a much more computationally intensive task available through the MIT-IBM Watson AI Lab. “I feel like that’s the kind of work we did on this project [first] project and in collaboration with all of our people on the IBM side, he was very helpful in figuring out how to make this change happen.” Moreover, Andreas’ group was able to implement multi-year projects on pre-training, reinforcement learning and calibration for reliable responses thanks to the computational resources and expertise of the MIT-IBM community.
Several other faculty members also found timely participation in the MIT-IBM Watson AI Lab to be very beneficial. “Having both the intellectual support and the opportunity to leverage some of the computational resources available at MIT-IBM has completely changed me and is extremely important to my research program,” says Yoon Kim — an associate professor in EECS, CSAIL and a researcher in the MIT-IBM Watson AI Lab — who has also noticed a change in the trajectory of his research field. Before joining MIT, Kim met his future collaborators during a postdoctoral fellowship at MIT-IBM, where he developed a neurosymbolic model; Currently, Kim’s team is developing methods to improve the capabilities and performance of the Huge Language Model (LLM).
One of the factors that led to his group’s success, which he points to, is a velvety research process involving intellectual partners. This enabled his MIT-IBM team to bid for the project, conduct large-scale experiments, identify bottlenecks, validate techniques, and adapt as necessary to develop state-of-the-art methods for potential real-world applications. “It’s an impetus for new ideas, and that’s what I think is unique about this relationship,” Kim says.
Pooling expertise
The nature of the MIT-IBM Watson AI Lab is such that it not only brings together artificial intelligence researchers to accelerate research, but also bridges work across disciplines. Lab researcher and MIT associate professor in EECS and CSAIL Justin Solomon describes his research group as growing up in the lab and collaboration as “crucial… from its inception to the present.” Solomon’s research team focuses on theoretically oriented geometric problems related to computer graphics, vision and machine learning.
Solomon credits the MIT-IBM collaboration with broadening his skills and the applications of his group’s work — a sentiment also shared by lab researchers Chuchu Fan, associate professor of aeronautics and astronautics and member of the Information and Decision Systems Laboratory, and Faez Ahmed, associate professor of mechanical engineering. “They [IBM] we’re able to translate some of these really messy engineering problems into the kind of math resources that our team can work on and close the loop,” Solomon says. For Solomon, that includes combining distinct AI models that have been trained on different datasets for distinct tasks. “I think these are really exciting spaces,” he says.
“I think these are career-starting projects [with the MIT-IBM Watson AI Lab] have largely shaped my own research agenda,” says Fan, whose research combines robotics, control theory, and safety-critical systems. Like Kim, Solomon, and Andreas, Fan and Ahmed began collaborative projects in their first year at MIT. Constraints and optimization govern the problems Fan and Ahmed tackle, and therefore require deep domain knowledge outside of artificial intelligence.
Collaborating with the MIT-IBM Watson AI Lab allowed Fan’s group to combine formal methods with natural language processing, which she says allowed the team to move from developing autoregressive task and motion planning for robots to creating LLM-based agents for travel planning, decision making, and verification. “This work was the first attempt to use LLM to translate any natural language in any form into a specification that a robot could understand and execute. This is something I am very proud of, but which was very difficult at the time,” says Fan. What’s more, through the collaborative investigation, her team was able to improve the LLM’s reasoning, which “would have been impossible without IBM’s support,” she says.
Through the lab, Faez Ahmed’s collaboration has facilitated the development of machine learning methods to accelerate the discovery and design of complex mechanical systems. Their Connections the work, for example, uses “generative optimization” to solve engineering problems in a way that is both data-driven and precise; recently apply multimodal data and LLM to computer-aided design. Ahmed says AI is often applied to problems that can already be solved but could benefit from increased speed and efficiency; however, challenges – such as mechanical connections that were considered “almost unsolvable” – are now within reach. “I think it’s definitely a feature [of our MIT-IBM team]” says Ahmed, praising the achievements of his MIT-IBM group, co-chaired by IBM’s Akash Srivastava and Dan Gutfreund.
What began as an initial collaboration between each MIT faculty member has evolved into an enduring intellectual relationship in which both parties are “excited about learning” and “student-oriented,” adds Ahmed. Taken together, the experiences of Jacob Andreas, Yoon Kim, Justin Solomon, Chuchu Fan, and Faez Ahmed demonstrate the impact that sustained, practical relationships between academia and industry can have on the formation of research groups and ambitious scientific exploration.