Kolmogorov-Arnold Networks Put Learnable Functions on Edges Instead of Fixed Node Activations
Standard multilayer perceptrons place weights on edges and fixed nonlinearities like ReLU at nodes. Kolmogorov-Arnold Networks, or KANs, instead build on the 1957 Kolmogorov-Arnold theorem, which expresses multivariate functions through univariate functions and addition. Each edge in a KAN carries a learnable univariate function, parameterized with B-splines plus a smooth basis, while nodes only sum their inputs. This design lets the learned edge functions be inspected individually, which the article presents as an interpretability advantage. The approach also allows grid refinement during training, which adds capacity without retraining from scratch.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.
Discussion (0)
Log in to join the discussion and vote.
Log in