【论文阅读】KAN: Kolmogorov–Arnold Networks（2024）

择要

Inspired by the Kolmogorov-Arnold representation theorem（体现定理）, we propose Kolmogorov-Arnold Networks (KANs) as promising alternatives（替换品） to Multi-Layer Perceptrons（多层感知机） (MLPs). While MLPs have fixed activation functions（固有的激活函数） on nodes（节点） (“neurons”), KANs have learnable activation functions（可学习的激活的函数） on edges（边沿） (“weights”). KANs have no linear weights at all（没有线性权重） – every weight parameter（权重参数） is replaced（代替） by a univariate function parametrized as a spline（参数化为样条的单变量函数）. We show that this seemingly simple change makes KANs outperform（优于） MLPs in terms of accuracy（精确性） and interpretability（可表明性）, on small-scale AI + Science tasks. For accuracy（在精度方面）, smaller KANs can achieve comparable or better accuracy（到达相称大概更好的精确度） than larger MLPs in function fitting tasks（拟合任务）. Theoretically and empirically（从理论和履历上看）, KANs possess faster neural scaling laws（具有更快的神经尺度规律） than MLPs. For interpretability（对于可表明性）, KANs can be intuitively visualized（直观地可