CS Seminar and CS460.org Research Talk 11/16: SPANDAN MADAN
11/16/2022 at 11am, McCormack M03-0430
Title: Role of data distributions in understanding the generalization behavior of biological and artificial intelligence
Abstract: Human vision generalizes to understand new, previously unseen scenes effortlessly. In stark contrast, computer vision models have been documented to struggle at generalizing to unseen object viewpoints, changes in scene lighting, small transformations including crops, rotations and are plagued by adversarial attacks. Understanding and correcting these behavioral artifacts lies at the heart of building robust models that can be deployed in the real world with confidence. In this talk, we will focus on how computer graphics serves as a bridge to investigate both biological and artificial intelligence. Showcasing the utility of controller 3D worlds created using graphics in understanding the generalization capabilities of ML models, designing new architectures which can generalize better, and studying how these behaviors differ from human vision.
Speaker bio: Spandan Madan is a senior graduate student at Harvard university housed in the departments of Computer Science and the Neuroscience. He is co-supervised by Professor Hanspeter Pfirster (SEAS) and Professor Gabriel kreiman (Boston children’s hospital), and his work focuses on the intersection of biological and artificial intelligence. In the past, he has been a student at IIT Delhi, and has gained industrial experience at Microsoft Research and Adobe Research. He is also a receipient of the Snap Research Scholarship, and the Winston Chen family graduate fellowship.