Kolmogorov-Arnold networks
组织者
Fedor Pavutnitskiy
演讲者
Pavel Tikhomirov
时间
2024年06月27日 15:15 至 16:15
地点
A3-3-301
线上
Zoom 928 682 9093
(BIMSA)
摘要
This talk examines the architecture of Kolmogorov-Arnold networks (KANs), covering their theoretical foundation, motivations, advantages, and limitations. We will discuss KANs’ interpretability and methods for model structure analysis. The presentation will address the trade-off between model size and accuracy in KANs compared to Multilayer Perceptrons (MLPs), using Pareto frontiers. Basic machine learning problems, such as the curse of dimensionality and catastrophic forgetting, will be reviewed in the context of KANs, highlighting training procedure and its connection to spline approximation. Applications of KANs in knot theory will be discussed, along with the general limitations and potential of KANs in various fields. The goal is to provide a concise overview of KANs’ theoretical and practical aspects.