Kolmogorov-Arnold Networks - Alternatives to Multi-Layer Perceptrons

Explore the innovative Kolmogorov-Arnold Networks (KANs) in this Google Algorithms Seminar TechTalk presented by Ziming Liu. Discover how KANs, inspired by the Kolmogorov-Arnold representation theorem, offer a promising alternative to Multi-Layer Perceptrons (MLPs) with learnable activation functions on edges instead of fixed activation functions on nodes. Learn about the advantages of KANs in terms of accuracy and interpretability, including their ability to achieve comparable or better results with smaller networks and faster neural scaling laws. Gain insights into how KANs can be visualized intuitively and interact with human users, making them valuable collaborators in (re)discovering mathematical and physical laws. Delve into the speaker's background as a PhD student at MIT & IAIFI, exploring his research interests at the intersection of AI and physics, and his contributions to various scientific fields.

KAN: Kolmogorov-Arnold Networks