KAN-MLP-Mixer: Hybrid Kolmogorov-Arnold Networks for IMU-based Human Activity Recognition

Forum topic · 小凯 · 2026-05-21

Summary

This paper investigates how Kolmogorov-Arnold Networks (KANs) can improve IMU-based human activity recognition (HAR). While KANs excel at learning complex functions on clean, low-dimensional data, they underperform on noisy real-world sensor data, where conventional MLPs remain more noise-tolerant and computationally efficient. The authors systematically explore placement of KAN modules within deep HAR networks and propose a hybrid KAN-MLP-Mixer architecture: a KAN-based input embedding layer, MLP layers for intermediate feature mixing, and a specialized LarctanKAN module for final classification. Evaluated on eight public HAR datasets, the hybrid model achieves an average 5.33% relative improvement in macro F1 score over pure-MLP models and outperforms standalone KAN and MLP baselines. Integrating the hybrid strategy into other state-of-the-art HAR architectures consistently boosts performance, showing that carefully orchestrated combinations of KANs and conventional neural components yield more robust and accurate HAR models for real-world wearable sensing.

Paper Overview

Research areas: cs.AI, eess.SP Authors: Mengxi Liu, Sizhen Bian, Vitor Fortes Published: 2026-05-21 arXiv: 2505.01255

Abstract

Kolmogorov-Arnold Networks (KANs) have demonstrated an exceptional ability to learn complex functions on clean, low-dimensional data but struggle to maintain performance on noisy and imperfect real-world datasets. In contrast, conventional multi-layer perceptrons (MLPs) are far more tolerant to noise and computationally efficient. Replacing all MLP components with KANs in HAR models often degrades accuracy and computation efficiency, highlighting an open challenge: how to combine KANs' precision with MLPs' noise robustness and efficiency.

To address this, the authors systematically explore various placements of KAN modules within deep HAR networks and propose a hybrid architecture that strategically synergizes the strengths of both paradigms:

KAN-based input embedding layer for expressive input transformation
MLP layers retained for intermediate feature mixing
Specialized LarctanKAN module for final activity classification

Results

Across eight public HAR datasets, the hybrid KAN-MLP model achieves an average macro F1 score relative improvement of 5.33% compared to the pure-MLP model, significantly outperforming standalone KAN and MLP baselines.
Integrating this hybrid strategy into other state-of-the-art HAR architectures consistently boosts their performance.

Conclusion

A carefully orchestrated combination of KAN, MLP, or other conventional neural components yields more robust and accurate HAR models for real-world wearable sensing environments.

--- *Source: arXiv:2505.01255*