The neural network architecture KAN has launched its 2.0 version, deepening its integration with scientific problems, especially in the field of classical physics research. Researchers can now customize their KAN2.0, integrating professional knowledge into the model to discover important concepts such as the Lagrangian in physical systems.
KAN2.0 allows researchers to tailor the model according to individual needs, using professional knowledge as auxiliary variables, providing new perspectives for the study of classical physics.
The new framework KAN2.0 is dedicated to addressing the inherent incompatibility between AI and science. It achieves the unification of AI and science through bidirectional synergy—integrating scientific knowledge into KAN and extracting scientific insights from KAN.
Three New Features of KAN2.0
MultKAN: Introduces multiplicative nodes into KAN, enhancing the model's expressive power.
kanpiler: A compiler that translates symbolic formulas into KAN, improving the model's practicality.
Tree Transformer: Converts the KAN2.0 architecture into a tree diagram, enhancing the model's interpretability.
KAN2.0's role in scientific discovery is mainly reflected in three aspects: identifying important features, revealing modular structures, and discovering symbolic formulas. These functions have been enhanced on the basis of the original KAN.
KAN2.0's interpretability is more universal, applicable to fields like chemistry and biology where symbolic equations are difficult to represent. Users can build modular structures into KAN2.0 and visually observe the modular structures through exchanges with MLP neurons.
The research team plans to apply KAN2.0 to larger-scale problems and extend it to other scientific disciplines beyond physics.
This research was jointly completed by five researchers from institutions such as MIT, Caltech, and MIT CSAIL, including three Chinese scholars. The first author of the paper, Liu Ziming, is a fourth-year PhD student at MIT, with research interests focused on the intersection of artificial intelligence and physics.
Paper link: https://arxiv.org/pdf/2408.10205
Project link: https://github.com/KindXiaoming/pykan