Look-Up Table Optimization for Kolmogorov-Arnold Networks on Microcontrollers

Kolmogorov-Arnold Networks (KANs) have emerged as interpretable alternatives to multilayer perceptrons, demonstrating strong performance in scientific computing. However, KAN inference requires evaluating computationally expensive basis functions (Jacobi polynomials or B-splines), creating a deployment barrier for resource-constrained microcontrollers (MCUs) without floating-point units (FPUs). We present LUT-KAN, a post-training quantization methodology that replaces basis-function evaluation with look-up table (LUT) operations and optional interpolation implemented with integer arithmetic. Benchmarks across 5,310 configurations on ARM Cortex-M3 demonstrate speedups of 2–25× depending on basis function type: Jacobi polynomials achieve 1.5–2.8× speedup while B-splines achieve 7–25× due to the high cost of Cox-de Boor recursion in the float baseline. LUT-KAN maintains low approximation error with as little as 68 bytes per network edge, enabling practical KAN deployment on MCUs with limited Flash and no FPU.