Research & Papers

Hybrid KAN-MLP model boosts activity recognition by 5.33%

New hybrid architecture combines KAN precision with MLP noise tolerance for wearable sensors.

Deep Dive

Kolmogorov-Arnold Networks (KANs) excel at learning complex functions on clean, low-dimensional data but struggle with noisy real-world sensor inputs. Conversely, traditional multi-layer perceptrons (MLPs) handle noise well but lack KANs' precision. A team of researchers led by Mengxi Liu systematically explored where to place KAN modules in deep HAR networks and developed KAN-MLP-Mixer, a hybrid that uses a KAN-based input embedding layer, retains MLP layers for intermediate feature mixing, and introduces a specialized LarctanKAN module for final classification.

Evaluated across eight public HAR datasets (IMU-based wearable sensing), the hybrid model delivered a 5.33% average macro F1 improvement over pure-MLP baselines, significantly outperforming standalone KAN and MLP architectures. The improvement was consistent when the hybrid strategy was integrated into other state-of-the-art HAR models. This demonstrates that a carefully orchestrated combination of KAN and MLP components yields more robust and accurate activity recognition for real-world noisy environments, advancing practical wearable AI applications.

Key Points
  • 5.33% average macro F1 improvement over pure-MLP baselines on 8 public HAR datasets
  • Novel LarctanKAN module used for final activity classification, replacing standard MLP layers
  • Hybrid outperforms both pure KAN and pure MLP baselines in noise-robustness and accuracy

Why It Matters

Improves real-world wearable activity recognition accuracy by combining KAN's precision with MLP's noise robustness.