Meet the Fellow: Ilia Sucholutsky
This entry is part of our Meet the Fellow blog series, which introduces and highlights Faculty Fellows who have recently joined CDS.
Meet CDS Faculty Fellow Ilia Sucholutsky, who will join the center this year. Sucholutsky recently completed his postdoctoral research in the Computational Cognitive Science Lab at Princeton University, supervised by Dr. Tom Griffiths. Prior to that, he earned his PhD in Statistics from the University of Waterloo, where he defended his thesis on “Learning From Almost No Data.”
Sucholutsky’s research explores the limits of learning from small datasets, investigating how much information machines and humans need to learn new tasks or categories. His work proposing that machines can learn to recognize more objects than the number of training examples that they are shown was widely acclaimed, and he and his team have since shown that humans can also learn in this surprisingly efficient way.
Recently, his research has expanded to examine learning as a form of communication between teachers and students, whether human or machine. In a recent NeurIPS publication, Sucholutsky has shown that the similarity between how a teacher and student represent the world, dictates their ability to efficiently communicate. Sucholutsky is leading a cross-disciplinary consortium of cognitive scientists, neuroscientists, and machine learning researchers to accelerate research on this kind of “representational alignment.” Together with this consortium, Sucholutsky has explored exciting applications of representational alignment to machine learning, natural language processing, robotics, human & machine perception, communication under uncertainty, value alignment, AI bias detection, and most recently, machine teaching.
“I’m excited to explore the notion of effective and efficient communication between humans and machines at NYU,” Sucholutsky said. “One thread I’m developing is flipping the typical script — instead of humans teaching machines, how can machines better teach humans?”
His current projects aim to create “machine tutors” or “machine thought partners” that can augment human capabilities. Sucholutsky’s team has found that many machine learning models, including large language models, tend to teach from their own perspective rather than adapting to the student’s way of understanding.
“We’ve shown recently that you can create student-centric teachers — teachers who take the time to understand how the student models the world, and then teach optimally to the student’s representations rather than their own,” Sucholutsky explained.
Sucholutsky was drawn to CDS by the concentration of researchers working at the intersection of human and machine learning. He looks forward to collaborating with faculty members like CDS Assistant Professor of Psychology and Data Science Brenden Lake, CDS Associate Professor of Linguistics and Data Science Tal Linzen, and others in the machine learning and computational cognitive science spheres.
By Stephen Thomas
