Evaluating Federated Kolmogorov-Arnold Networks on Non-IID Data

TL;DR

This paper compares Federated Kolmogorov-Arnold Networks (F-KANs) with MLPs on non-IID data, showing F-KANs can match MLP accuracy faster with moderate additional computation.

Contribution

It provides the first comparative evaluation of F-KANs and MLPs in federated learning on non-IID data, highlighting efficiency advantages of F-KANs.

Findings

F-KANs achieve similar accuracy to MLPs in fewer rounds.

Spline-KANs outperform Radial Basis Function KANs.

F-KANs require moderate additional computation time.

Abstract

Federated Kolmogorov-Arnold Networks (F-KANs) have already been proposed, but their assessment is at an initial stage. We present a comparison between KANs (using B-splines and Radial Basis Functions as activation functions) and Multi- Layer Perceptrons (MLPs) with a similar number of parameters for 100 rounds of federated learning in the MNIST classification task using non-IID partitions with 100 clients. After 15 trials for each model, we show that the best accuracies achieved by MLPs can be achieved by Spline-KANs in half of the time (in rounds), with just a moderate increase in computing time.