| The growth of edge computing and tiny machine learning (TinyML) applications demands neural network architectures that are accurate and highly efficient. Though Multi-Layer Perceptrons (MLPs) remain the traditional architecture, recently Kolmogorov-Arnold Networks (KANs) have been proposed as an alternative because of the promising gains in trade-offs between accuracy, parameter efficiency, and interpretability. Despite their promise, how KANs behave on resource-constrained microcontroller units (MCUs) remains unclear. To address this gap, the present work conducts a comprehensive benchmarking of KANs against equivalent MLPs trained, ported, and quantized under real-world edge conditions. This work benchmarks both architectures on four fundamental machine learning (ML) tasks: (1) Symbolic regression and function approximation using the Feynman benchmark (KAN-only); (2) Time series forecasting on climate and energy consumption datasets; (3) General tabular data tasks using the UCI regression and classification datasets; and (4) Image classification on standard image classification datasets. This benchmark analyses the performance, memory footprint, and latency of both KANs and MLPs on edge hardware, including the Arduino BLE 33 Sense (ARM Cortex-M4), Raspberry Pi Pico (ARM Cortex-M0+), and ESP32 (Tensilica Xtensa LX6). The benchmark analysis shows that MLPs are significantly more parameter-efficient and faster for tabular tasks, while KANs can offer higher accuracy and better robustness to INT8 quantization. Surprisingly, for image and time-series tasks, KANs demonstrate lower inference latency despite their larger model size. The results highlight a critical, task-dependent trade-off, demonstrating that the choice between KAN and MLP for TinyML is highly nuanced. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.