I'm a researcher working at the intersection of AI and neuroscience. Currently I do research at Mila - Quebec AI Institute, where I am advised by Prof. Shahab Bakhtiari. My research interest lies in representation learning, particularly, how AI systems, humans, and animals learn representations that capture the structure and dynamics of their environment, and enable interactions with the world such as control, planning, and theory of mind. I'm interested in how learning objectives, architectures, and data selection shape these representations, and the connection between good representations and performance in real-world interactions.
Previously, I graduated with a physics major from the University of Pennsylvania, working with Profs. Konrad Kording and Osbert Bastani.
How do Vision Transformers understand scenes with multiple objects without confusing their features? We formalized feature-to-object "binding" with an information-theoretic approach and devised a probing method to quantify binding information in the representations of any artificial model. Our work has important implications for object-centric perception in visual reasoning and embodied AI. Published at ICML 2026.
Feel free to reach out to me through my email: leoh@sas.upenn.edu