Gated QKAN-FWP: A Scalable Quantum-Inspired Sequence Learning Framework
A new research paper proposes gated QKAN-FWP, a fast-weight framework integrating Fast Weight Programmers (FWPs) with Quantum-inspired Kolmogorov-Arnold Networks (QKAN). It uses single-qubit data re-uploading circuits (DARUAN) as learnable nonlinear activations, avoiding multi-qubit architectures that are hard to scale on NISQ devices. The framework introduces a scalar-gated fast-weight update rule with theoretical guarantees on adaptive memory, geometric boundedness, and parallelizable gradients. Evaluated on time-series benchmarks and MiniGrid reinforcement learning tasks, it shows improved scalability and performance over existing quantum FWPs. The paper is available on arXiv under ID 2605.06734.
Key facts
- Proposes gated QKAN-FWP integrating FWP with Quantum-inspired Kolmogorov-Arnold Networks
- Uses single-qubit data re-uploading circuits (DARUAN) as activation functions
- Introduces scalar-gated fast-weight update rule
- Provides theoretical analysis of adaptive memory kernel, geometric boundedness, and parallelizable gradient paths
- Evaluated on time-series benchmarks and MiniGrid reinforcement learning tasks
- Aims to improve scalability on NISQ devices and reduce classical simulation cost
- arXiv ID: 2605.06734
- Publication type: cross
Entities
Institutions
- arXiv