QKAN: Quantum Kolmogorov-Arnold Networks
Preprint
Abstract
The potential of learning models in quantum hardware remains an open question. Yet, the field of quantum machine learning persistently explores how these models can take advantage of quantum implementations. Recently, a new neural network architecture, called Kolmogorov-Arnold Networks (KAN), has emerged, inspired by the compositional structure of the Kolmogorov-Arnold representation theorem. In this work, we design a quantum version of KAN called QKAN. Our QKAN exploits powerful quantum linear algebra tools, including quantum singular value transformation, to apply parameterized activation functions on the edges of the network. QKAN is based on block-encodings, making it inherently suitable for direct quantum input. Furthermore, we analyze its asymptotic complexity, building recursively from a single layer to an end-to-end neural architecture. The gate complexity of QKAN scales linearly with the cost of constructing block-encodings for input and weights, suggesting broad applicability in tasks with high-dimensional input. QKAN serves as a trainable quantum machine learning model by combining parameterized quantum circuits with established quantum subroutines. Lastly, we propose a multivariate state preparation strategy based on the construction of the QKAN architecture.Keywords
All Related Versions
This publication has 75 references indexed in Scilit:
- Kolmogorov–Arnold-Informed neural network: A physics-informed deep learning framework for solving forward and inverse problems based on Kolmogorov–Arnold NetworksComputer Methods in Applied Mechanics and Engineering, 2024
- On the interpretability of quantum neural networksQuantum Machine Intelligence, 2024
- Machine Learning with Quantum Computers2021
- On a Constructive Proof of Kolmogorov’s Superposition TheoremConstructive Approximation, 2009
- Representation of continuous functions of three variables by the superposition of continuous functions of two variables2009
- On functions of three variables2009
- Rademacher and Gaussian Complexities: Risk Bounds and Structural Results2001
- Approximation by superpositions of a sigmoidal functionMathematics of Control, Signals, and Systems, 1989
- Multilayer feedforward networks are universal approximatorsNeural Networks, 1989
- A logical calculus of the ideas immanent in nervous activityBulletin of Mathematical Biology, 1943