KAN-AE with Non-Linearity Score and Symbolic Regression for Energy-Efficient Channel Coding

TL;DR

This paper presents a novel approach combining Kolmogorov-Arnold network autoencoders with symbolic regression and a non-linearity score to create energy-efficient, low-complexity channel coding solutions that maintain competitive performance.

Contribution

It introduces a method to convert KAN-AEs into symbolic expressions with a non-linearity score, enhancing energy efficiency and implementation simplicity in channel coding.

Findings

KAN-AEs achieve competitive BLER performance.

The proposed method reduces energy consumption by 1.38 times compared to MLP-AEs.

Symbolic regression with non-linearity scoring improves model simplicity and efficiency.

Abstract

In this paper, we investigate Kolmogorov-Arnold network-based autoencoders (KAN-AEs) with symbolic regression (SR) for energy-efficient channel coding. By using SR, we convert KAN-AEs into symbolic expressions, which enables low-complexity implementation and improved energy efficiency at the radios. To further enhance the efficiency, we introduce a new non-linearity score term in the SR process to help select lower-complexity equations when possible. Through numerical simulations, we demonstrate that KAN-AEs achieve competitive BLER performance while improving energy efficiency when paired with SR. We score the energy efficiency of a KAN-AE implementation using the proposed non-linearity metric and compare it to a multi-layer perceptron-based autoencoder (MLP-AE). Our experiment shows that the KAN-AE paired with SR uses 1.38 times less energy than the MLP-AE, supporting that KAN-AEs are a promising choice for energy-efficient deep learning-based channel coding.