An asymmetric lightweight multi-task network based on KAN convolution

Fanyun Meng

doi:10.1007/s00500-025-10915-2

Return

An asymmetric lightweight multi-task network based on KAN convolution

2025-09-26

PRE

DOI：10.1007/s00500-025-10915-2

Original

Original request for help

Save

Abstract

En 中文

Multi-task learning has emerged as a critical paradigm in computer vision, enabling simultaneous processing of related tasks such as semantic segmentation and depth estimation. However, existing multi-task networks suffer from severe parameter inefficiency, with studies showing that up to 60–70% of parameters contribute minimally to model performance, leading to computational waste and suboptimal results. This paper presents LAKNet (Lightweight Asymmetric KAN Network), a novel multi-task architecture that fundamentally addresses parameter underutilization through three key innovations: (1) KAN (Kolmogorov-Arnold Network) convolution modules that employ pre-activation mechanisms to enhance parameter efficiency by 35–40% compared to traditional convolutions, (2) an asymmetric encoder-decoder structure that reduces parameters by 50% while maintaining performance through strategic deep-shallow feature fusion, and (3) a multi-scale channel attention mechanism that compensates for the reduced parameter count by enriching feature representations. Extensive experiments on the Cityscapes dataset demonstrate that LAKNet achieves state-of-the-art performance with 58.90% mIoU for semantic segmentation and 0.0182 absolute error for depth estimation, using only 3.86 × 107 parameters—outperforming MTAN (51.91% mIoU with 4.12 × 107 parameters) and other baseline methods. The theoretical analysis reveals that KAN convolution's pre-activation strategy enables each parameter to contribute more effectively to feature learning, addressing the fundamental inefficiency of post-activation architectures. Our code is available at https://github.com/SyqxhSt/Net-LAK-MTL

Keywords:

Multi-task learning

Kolmogorov-Arnold networks

Parameter efficiency

Asymmetric architecture

Channel attention