Latest News

Researchers from the German Research Center for Artificial Intelligence (DFKI) and ETH Zurich have introduced a novel neural architecture for sensor-based Human Activity Recognition (HAR), spotlighting the potential of Kolmogorov-Arnold Networks (KANs) for low-power, high-accuracy AI. Their paper, “Initial Investigation of Kolmogorov-Arnold Networks (KANs) as Feature Extractors for IMU Based Human Activity Recognition”, offers a new approach to building sustainable AI models optimized for time-series data from inertial sensors.

This work aligns closely with the goals of the SustainML project, which seeks to drive lifecycle-aware, application-specific model development for reducing AI’s environmental impact.

Rethinking Feature Extraction for Time-Series AI

Unlike traditional CNN-based feature extractors, KANs replace weight-based aggregation and static nonlinearities with spline-learned nonlinear transformations on input edges. These are better suited for real-valued, time-dependent IMU signals, the kind commonly found in HAR applications.

The paper explores four KAN-based architectures, 1L-KAN, 2L-KAN, RL-KAN, and PL-KAN—and demonstrates their effectiveness across various datasets. The key idea: KANs can learn signal patterns that are functionally expressive and resource-efficient, much like expert-crafted features from the pre-deep learning era, but in a fully trainable pipeline.

Significant Gains with Smaller Models

Experiments on four benchmark HAR datasets, WISDM, MotionSense, MM-Fit, and PAMAP2, demonstrated that KAN-based architectures:

• Achieve comparable or superior performance to CNNs, with 10–20× fewer parameters on simpler tasks.
• Outperform CNN baselines by 3–5% in macro F1 score on complex activity datasets.
• Offer smaller memory and energy footprints, which is critical for mobile, edge, or wearable deployment.

Notably, the RL-KAN FE architecture, which combines both channel-wise and cross-channel spline-based encoders, delivered the best overall results, suggesting that KANs can excel in learning both individual and combined sensor dynamics.

Building a Path Toward Sustainable AI on the Edge

This research directly supports SustainML’s objective of embedding sustainability into every stage of AI system design:

• Energy-Efficiency by Architecture
  KANs reduce the demand for excessive GPU computation, aligning well with hardware constraints on embedded and wearable devices.
• Smaller Models, Broader Applicability
  By shifting away from large, pre-trained CNN backbones, developers can deploy leaner AI systems in real-world applications such as mobile health, smart homes, and industrial monitoring.
• Resource-Conscious Innovation
  The study opens a new front for sustainability in AI: rather than sacrificing model expressiveness, KANs deliver a better function-fit per parameter, challenging the norm that accuracy must come with scale.

A Model for SustainML’s Vision

By demonstrating how domain-specific architecture choices can significantly reduce computational burden while preserving, or even improving, performance, DFKI’s work strengthens the case for intelligent, sustainable ML design.

“KANs offer an alternative that is both principled and practical: fewer parameters, more relevant features, and better alignment with the data’s underlying structure,” the authors argue.

This research is a step forward in the SustainML roadmap toward greener, smarter AI, especially for ubiquitous sensing environments where sustainability is not optional, but essential.