KAN-MLP-Mixer: A comprehensive investigation of the usage of Kolmogorov-Arnold Networks (KANs) for improving IMU-based Human Activity Recognition

Researchers propose KAN-MLP-Mixer, a hybrid neural network architecture that combines Kolmogorov-Arnold Networks (KANs) with traditional MLPs for human activity recognition from IMU sensors. The model achieves 5.33% improvement over pure-MLP baselines by leveraging KANs' precision in input embedding and classification while retaining MLPs' noise robustness for intermediate processing.

This research addresses a fundamental challenge in modern machine learning: balancing mathematical elegance with practical robustness. Kolmogorov-Arnold Networks represent a theoretical advancement in function approximation, yet their application to real-world sensor data reveals a critical limitation—vulnerability to noise inherent in wearable devices. The researchers' systematic exploration of architectural placement demonstrates that neural network design involves crucial trade-offs between theoretical capabilities and empirical performance.

The hybrid approach reflects broader trends in deep learning where practitioners increasingly recognize that no single component dominates all scenarios. By strategically positioning KANs for input transformation and final classification while preserving MLPs for intermediate feature processing, the team creates a framework that exploits each method's comparative advantage. This modular thinking enables the architecture to maintain computational efficiency while improving accuracy on real-world IMU datasets—a critical consideration for edge devices in wearable applications.

For practitioners developing activity recognition systems, this work validates the value of architectural experimentation beyond purely theoretical frameworks. The consistent performance improvements across eight public datasets suggest the hybrid strategy generalizes across different sensor configurations and activity types. The integration into existing state-of-the-art architectures indicates practitioners can retrofit this approach without complete redesigns.

Looking forward, the methodology extends beyond activity recognition to other sensor-based applications where noise tolerance and precision both matter. This research may catalyze broader adoption of hybrid architectures in edge AI and wearable computing, where computational constraints and real-world imperfections demand pragmatic rather than purist solutions.

• →Hybrid KAN-MLP architecture achieves 5.33% F1-score improvement over pure-MLP models on human activity recognition tasks
• →Strategic placement of KANs in input embedding and classification layers outperforms replacing all MLP components with KANs
• →The approach balances KANs' mathematical precision with MLPs' superior noise robustness for real-world wearable sensor data
• →Hybrid strategy consistently enhances state-of-the-art HAR architectures, suggesting broad applicability beyond baseline comparisons
• →Research validates modular neural architecture design as more effective than theoretical purity for practical machine learning applications