Compact Feedback-Enhanced Quantum Fast Weight Programmer for Time Series Forecasting
Year: 2026
Type: conference-paper
Abstract: Quantum recurrent models are principled for temporal learning, yet their reliance on backpropagation through time and deep unrolling is poorly matched to Noisy Intermediate-Scale Quantum constraints. Quantum Fast Weight Programming alleviates this depth bottleneck via shallow, online parameter adaptation, but remains sensitive to noise-induced variance and gradient degradation, especially when the programmer has limited feedback about the current fast-weight state. This work tries to overcome these challenges by introducing the Compact Feedback-Enhanced Quantum Fast Weight Programmer, a hybrid architecture that strengthens quantum fast-weight programming with an explicit state-aware controller while preserving shallow, hardware-efficient quantum execution. To maintain parameter efficiency, the controller generates rank-1 parameter increments via the outer product of layer-wise and qubit-wise modulation vectors, avoiding full-matrix generation while supporting expressive adaptation. The proposed approach is assessed on three real-world forecasting benchmarks spanning disparate fields and dynamics, consistently improving predictive accuracy over both quantum fast weight programming approaches and quantum deep recurrent neural networks, while requiring also substantially less training time than quantum baselines.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7222924925