The neural network architecture KAN has ushered in a major upgrade - KAN 2.0 version. This update significantly enhances KAN's application capabilities in scientific research, especially in the field of classical physics. Researchers are able to customize KAN 2.0 models and incorporate expertise to better explore physical systems, such as identifying key concepts such as the Lagrangian. This marks another leap in the application of AI in scientific research and provides a new way to solve the inherent incompatibility problem between AI and science.
The neural network architecture KAN has ushered in its version 2.0. This update makes KAN more deeply integrated with scientific issues, especially in the field of classical physics research. Researchers can now customize their own KAN2.0 and incorporate expertise into the model to discover important concepts such as the Lagrangian in physical systems.
KAN2.0 allows researchers to customize models according to personal needs, using professional knowledge as an auxiliary variable, providing a new perspective for the study of classical physics.
The new framework KAN2.0 is dedicated to solving the inherent incompatibility problem between AI and science. It unifies AI and science through two-way synergy - integrating scientific knowledge into KAN and extracting scientific insights from KAN.
Three new functions of KAN2.0
MultKAN: KAN that introduces multiplication nodes enhances the expression ability of the model.
kanpiler: A compiler that compiles symbolic formulas into KAN, improving the practicality of the model.
Tree converter: Converts the KAN2.0 architecture into a tree diagram, enhancing the interpretability of the model.
The role of KAN2.0 in scientific discovery is mainly reflected in three aspects: identifying important features, revealing module structures, and discovering symbolic formulas. These features are enhanced over the original KAN.
The interpretability of KAN2.0 is more general and suitable for fields such as chemistry and biology that are difficult to express by symbolic equations. Users can build modular structures into KAN2.0 and visually see the modular structures by exchanging with MLP neurons.
The research team plans to apply KAN2.0 to larger-scale problems and expand it to other scientific disciplines beyond physics.
This research was jointly completed by five researchers from MIT, California Institute of Technology, MIT CSAIL and other institutions, including three Chinese scholars. Liu Ziming, the first author of the paper, is a fourth-year doctoral student at MIT. His research interests focus on the intersection of artificial intelligence and physics.
Paper address: https://arxiv.org/pdf/2408.10205
Project address: https://github.com/KindXiaoming/pykan
The release of KAN 2.0 provides a powerful new tool for scientific research with promising applications in interdisciplinary fields. In the future, we can expect KAN 2.0 to make breakthroughs in more scientific fields and promote the accelerated development of scientific discoveries. The source code of this project has been open sourced, and everyone is welcome to contribute.