Speaker
Description
Variational Quantum Recurrent Neural Networks (QRNNs) offer a promising approach to time series forecasting on quantum hardware. However, despite encouraging simulation results, physical deployment faces three critical challenges: trainability issues caused by barren plateaus and hardware noise; high execution overhead when computing analytical gradients via parameter-shift rules; and mid-circuit measurements that introduce severe latency and thermal decoherence. While distributed multi-QPU systems with fast, high-fidelity gates and measurements could alleviate these bottlenecks, such hardware environments remain largely unavailable.
To circumvent these limitations on current hardware, we propose QPU partitioning as an immediate, single-chip solution for training QRNNs. By executing independent logical circuits in parallel on a single physical QPU, each one in a subset of its qubits, we exploit the uniform structure of parameter-shift gradient evaluations, where parallel circuits share similar execution times and allow trivial scheduling. We evaluate this approach by comparing gradient accuracy and execution times across ideal density-matrix simulations, unpartitioned QPU deployments, and partitioned QPU runs with and without dynamical decoupling. Using distance metrics to compare the output probability distributions, we demonstrate that QPU partitioning is both feasible on existing quantum devices and significantly reduces training time for variational QRNNs.