摘要
Inspired by the Kolmogorov-Arnold representation theorem(表示定理), we propose Kolmogorov-Arnold Networks (KANs) as promising alternatives(替代品) 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(边缘) (“weights”). KANs have no linear weights at all(没有线性权重) – every weight parameter(权重参数) is replaced(取代) 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(准确性) and interpretability(可解释性), on small-scale AI + Science tasks. For accuracy(在精度方面), smaller KANs can achieve comparable or better accuracy(达到相当或者更好的准确度ÿ