Abstract:In this paper, we propose to use Sinc interpolation in the context of Kolmogorov-Arnold Networks, neural networks with learnable activation functions, which recently gained attention as alternatives to multilayer perceptron. Many different function representations have already been tried, but we show that Sinc interpolation proposes a viable alternative, since it is known in numerical analysis to represent well both smooth functions and functions with singularities. This is important not only for function approximation but also for the solutions of partial differential equations with physics-informed neural networks. Through a series of experiments, we show that SincKANs provide better results in almost all of the examples we have considered.

What problem does this paper attempt to address?

### Problems the Paper Attempts to Solve The paper proposes a new network architecture—Sinc Kolmogorov-Arnold Network (SincKAN), aiming to address the following issues: 1. **Ability to Handle Singularities**: - SincKAN utilizes Sinc function interpolation instead of traditional cubic spline interpolation to improve the activation function. The Sinc function performs excellently in handling problems with singularities, allowing for better approximation of functions with singular points. 2. **Mitigating Spectral Bias**: - Physics-Informed Neural Networks (PINNs) face the issue of spectral bias when solving partial differential equations. SincKAN alleviates this spectral bias by handling singularities, thereby enhancing the performance of PINNs. 3. **Enhancing PINNs Performance**: - Through a series of experimental validations, SincKAN demonstrates outstanding performance in various benchmarks, especially in handling boundary layer problems and problems with singularities. Compared to other networks (such as MLP, KAN, and ChebyKAN), SincKAN shows better accuracy and generalization ability. 4. **Optimizing Step Size Selection**: - To select the optimal step size \( h \), the paper proposes a multi-step interpolation method. Experimental results indicate that this method is the main reason SincKAN can better approximate complex smooth functions. 5. **Coordinate Transformation and Normalization**: - The paper introduces normalization transformations to select appropriate coordinate transformations and proposes skip connections with learnable linear functions to meet decay conditions. These improvements make SincKAN more robust and easier to implement in practical applications. In summary, the paper aims to improve the performance of PINNs through SincKAN, particularly for problems with singularities and boundary layer issues, to enhance its application effectiveness in scientific computing.

Sinc Kolmogorov-Arnold Network and Its Applications on Physics-informed Neural Networks

SineKAN: Kolmogorov-Arnold Networks Using Sinusoidal Activation Functions

Kolmogorov Arnold Informed neural network: A physics-informed deep learning framework for solving forward and inverse problems based on Kolmogorov Arnold Networks

Kolmogorov Arnold Informed neural network: A physics-informed deep learning framework for solving PDEs based on Kolmogorov Arnold Networks

Chebyshev Polynomial-Based Kolmogorov-Arnold Networks: An Efficient Architecture for Nonlinear Function Approximation

KKANs: Kurkova-Kolmogorov-Arnold Networks and Their Learning Dynamics

KAN: Kolmogorov-Arnold Networks

NEW DATA ANALYSIS ALGORITHMS BASED ON SPLINE VERSIONS OF KOLMOGOROV ARNOLD NETWORKS

Smooth Kolmogorov Arnold networks enabling structural knowledge representation

Physics Informed Kolmogorov-Arnold Neural Networks for Dynamical Analysis via Efficent-KAN and WAV-KAN

Physics-informed Kolmogorov-Arnold Network with Chebyshev Polynomials for Fluid Mechanics

Kolmogorov--Arnold networks in molecular dynamics

A Survey on Kolmogorov-Arnold Network

A comprehensive and FAIR comparison between MLP and KAN representations for differential equations and operator networks

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

Deep Learning Alternatives of the Kolmogorov Superposition Theorem

Convolutional Kolmogorov-Arnold Networks

MonoKAN: Certified Monotonic Kolmogorov-Arnold Network

P1-KAN an effective Kolmogorov Arnold Network for function approximation

LArctan-SKAN: Simple and Efficient Single-Parameterized Kolmogorov-Arnold Networks using Learnable Trigonometric Function

Generalization Bounds and Model Complexity for Kolmogorov-Arnold Networks