LTBs-KAN: Enabling Linear-Time B-Spline Computation for KAN Networks

Introduction: A Solution Emerges for KAN's Speed Dilemma

Since Kolmogorov-Arnold Networks (KAN) burst onto the scene in 2024, this entirely new neural network architecture has rapidly become a research hotspot in deep learning, thanks to its superior interpretability and expressive power compared to multilayer perceptrons (MLPs). However, KAN's critical weakness has been equally apparent — due to the recursive nature of B-spline function computation, its training and inference speeds are far slower than traditional MLPs, severely limiting the expansion of practical application scenarios.

Recently, a paper published on arXiv (arXiv:2604.22034) introduced a novel architecture called "LTBs-KAN" (Linear-Time B-splines Kolmogorov-Arnold Network). By redesigning the computation method for B-spline basis functions and reducing time complexity to a linear level, it clears a critical obstacle for large-scale deployment of KAN networks.

Core Technology: B-Spline Innovation from Recursive to Linear

The Root Cause of KAN's Speed Bottleneck

The core idea of traditional KAN architecture originates from the Kolmogorov-Arnold representation theorem, replacing fixed activation functions in MLPs with learnable one-dimensional functions (typically implemented with B-splines). This design endows KAN with powerful function approximation capabilities and strong interpretability, but also introduces computational cost issues.

Standard B-spline function computation relies on the de Boor recursion algorithm, whose time complexity is closely tied to the spline order. As network depth increases and nodes multiply, the overhead from recursive computation grows multiplicatively, causing KAN's training speed to be several times or even tens of times slower than equivalently sized MLPs.

LTBs-KAN's Solution

The core innovation of LTBs-KAN lies in proposing a "linear-time B-spline" computation method that fundamentally changes the evaluation strategy for B-spline basis functions. The research team restructured the base-spline computation workflow, avoiding the extensive redundant operations in traditional recursive structures, reducing the evaluation time complexity of each spline basis function to O(1) level, and achieving linear time complexity for overall network forward propagation.

This improvement is not an approximation method that sacrifices accuracy. Instead, through mathematically equivalent transformations, it achieves an efficiency leap while preserving the complete mathematical properties of B-splines. This means LTBs-KAN theoretically possesses exactly the same expressive power and interpretability as the original KAN.

In-Depth Analysis: The Significance and Impact of LTBs-KAN

Bridging the Efficiency Gap Between KAN and MLP

Since KAN's inception, academic assessment has consistently shown a split between "theoretically promising" and "practically limited." Although extensive experiments have demonstrated KAN's excellent performance in low-dimensional function fitting, scientific computing, symbolic regression, and other tasks, its speed disadvantage has made it difficult to compete with MLPs in large-scale tasks such as computer vision and natural language processing. The emergence of LTBs-KAN is expected to significantly narrow this gap, making KAN a viable alternative to MLPs across a broader range of application scenarios.

Accelerating the Evolution of the KAN Ecosystem

Since 2024, improvement research surrounding KAN has proliferated, with multiple variants including EfficientKAN, FastKAN, and FourierKAN emerging in succession, forming a rapidly growing research ecosystem. LTBs-KAN focuses on the most fundamental B-spline computational efficiency problem, taking a "solving the problem at its root" approach. Its technical solution could potentially be adopted and integrated by other KAN variants, thereby driving the maturation of the entire KAN ecosystem.

New Momentum for Explainable AI

At a time when AI interpretability is receiving increasing attention, KAN has garnered significant interest due to its inherently interpretable structure. LTBs-KAN solves the efficiency problem without sacrificing interpretability, which holds particularly important value for fields with rigid demands for "explainable AI," such as medical diagnosis, financial risk management, and scientific discovery. When KAN's speed is no longer an obstacle, practitioners in these fields will have a new tool that combines both performance and transparency.

Outlook: How Far Is KAN from the Mainstream?

Although LTBs-KAN has achieved an important breakthrough on the efficiency front, KAN still faces multiple challenges before it can become a mainstream deep learning architecture. First, KAN's performance on ultra-large-scale data and ultra-high-dimensional tasks still requires more empirical validation. Second, the engineering ecosystem surrounding KAN (framework support, hardware optimization, pre-trained models) remains in its early stages. Additionally, methods for integrating KAN with mainstream architectures such as Transformers require further exploration.

Nevertheless, the trend represented by LTBs-KAN is clear — KAN is rapidly transitioning from an "interesting theoretical innovation" to a "usable practical tool." As efficiency bottlenecks are progressively overcome, KAN is expected to achieve substantive replacement of MLPs in specific vertical domains within the next one to two years, becoming an indispensable member of the deep learning toolbox.

For AI researchers and engineers, now may be the best time to start paying serious attention to the KAN technology roadmap.