Abstract:Kolmogorov-Arnold Networks(KANs), as a theoretically efficient neural network architecture, have garnered attention for their potential in capturing complex patterns. However, their application in computer vision remains relatively unexplored. This study first analyzes the potential of KAN in computer vision tasks, evaluating the performance of KAN and its convolutional variants in image classification and semantic segmentation. The focus is placed on examining their characteristics across varying data scales and noise levels. Results indicate that while KAN exhibits stronger fitting capabilities, it is highly sensitive to noise, limiting its robustness. To address this challenge, we propose a smoothness regularization method and introduce a Segment Deactivation technique. Both approaches enhance KAN's stability and generalization, demonstrating its potential in handling complex visual data tasks.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to evaluate the performance of Kolmogorov - Arnold Networks (KAN) and their convolutional variants in computer vision tasks, especially in image classification and semantic segmentation tasks. The focus of the research is to examine the characteristics of these models under different data scales and noise levels. Specifically, the paper focuses on the following aspects: 1. **Evaluating the potential of KAN in visual tasks**: Researchers first analyzed the potential applications of KAN in computer vision tasks and experimentally evaluated the performance of KAN and convolutional KAN (CKAN) in image classification and semantic segmentation tasks. 2. **Exploring the limitations of KAN**: The experimental results show that although KAN has a strong fitting ability, it is highly sensitive to noise, which limits its robustness in practical visual applications. 3. **Proposing improvement methods**: In order to overcome the problem of KAN's sensitivity to noise, the researchers proposed the smoothing regularization method and the smoothing deactivation technique (Segment Deactivation). These two methods aim to enhance the stability and generalization ability of KAN, enabling it to better handle complex visual data tasks. ### Main contributions 1. **Evaluating the performance of KAN in visual tasks**: By systematically evaluating the performance of KAN and its variants in key computer vision tasks (such as image classification, object detection, and segmentation), their advantages and limitations are identified. 2. **Smoothing regularization method**: A smoothing regularization method is proposed. By constraining the rate of change of the spline function, the excessive fluctuation of model parameters is reduced, thereby improving the stability of the model and enhancing its robustness in noisy environments. 3. **Smoothing deactivation technique**: A new technique - smoothing deactivation is introduced. This technique simplifies certain spline segments into linear functions with a certain probability during the training process, thereby effectively enhancing the robustness of the training process, simplifying the complexity of the model, and improving the performance of KAN in high - dimensional visual tasks. ### Experimental design - **Preliminary evaluation**: Preliminary experiments were carried out on the CIFAR - 100 and PASCAL VOC2012 datasets, and the performance of KAN and CKAN was compared with that of baseline models (such as MobileNet and UNet). - **Influence of data scale**: By changing the size of the dataset (from 20% to 100%), the change in the fitting ability of KAN and CKAN with the increase in data volume was studied. - **Noise robustness**: The performance of the model was evaluated under different levels of label noise to verify the sensitivity of KAN and CKAN to noise. - **Influence of regularization**: By introducing L1 regularization, the influence of regularization on the generalization ability of KAN and CKAN was studied. - **Effect of improvement methods**: The effectiveness of the smoothing regularization method and the smoothing deactivation technique was tested on the CIFAR - 10 and CIFAR - 100 datasets. The results show that these methods significantly improve the performance and robustness of KAN. In conclusion, this paper evaluates the potential and limitations of KAN in computer vision tasks through a series of experiments and proposes effective improvement methods, providing theoretical and technical support for the application of KAN in complex visual tasks.

Can KAN Work? Exploring the Potential of Kolmogorov-Arnold Networks in Computer Vision

KANs for Computer Vision: An Experimental Study

Suitability of KANs for Computer Vision: A preliminary investigation

Demonstrating the Efficacy of Kolmogorov-Arnold Networks in Vision Tasks

A Comprehensive Survey on Kolmogorov Arnold Networks (KAN)

Kolmogorov-Arnold Convolutions: Design Principles and Empirical Studies

A Survey on Kolmogorov-Arnold Network

Residual Kolmogorov-Arnold Network for Enhanced Deep Learning

Convolutional Kolmogorov-Arnold Networks

Kolmogorov-Arnold Network Autoencoders

KAN: Kolmogorov-Arnold Networks

A preliminary study on continual learning in computer vision using Kolmogorov-Arnold Networks

A Revolutionary Neural Network Architecture with Interpretability and Flexibility Based on Kolmogorov–Arnold for Solar Radiation and Temperature Forecasting

Kolmogorov-Arnold Networks (KAN) for Time Series Classification and Robust Analysis

On the expressiveness and spectral bias of KANs

How to Learn More? Exploring Kolmogorov–Arnold Networks for Hyperspectral Image Classification

Kolmogorov-Arnold Network for Satellite Image Classification in Remote Sensing

Generalization Bounds and Model Complexity for Kolmogorov-Arnold Networks

Reimagining Linear Probing: Kolmogorov-Arnold Networks in Transfer Learning

Exploring the Limitations of Kolmogorov-Arnold Networks in Classification: Insights to Software Training and Hardware Implementation

KANICE: Kolmogorov-Arnold Networks with Interactive Convolutional Elements