[Paper Review] Practical application-specific advantage through hybrid quantum computing

Michael Perelshtein, Asel Sagingalieva+8 more|arXiv (Cornell University)|May 10, 2022

Quantum Computing Algorithms and ArchitectureComputer Science21 citations

TL;DR

The paper presents a memory-centric hybrid quantum cloud (QMware) integrating quantum and classical resources to solve optimization, machine learning, and simulation tasks, showing practical advantages over classical methods for specific problems.

ABSTRACT

Quantum computing promises to tackle technological and industrial problems insurmountable for classical computers. However, today's quantum computers still have limited demonstrable functionality, and it is expected that scaling up to millions of qubits is required for them to live up to this touted promise. The feasible route in achieving practical quantum advantage goals is to implement a hybrid operational mode that realizes the cohesion of quantum and classical computers. Here we present a hybrid quantum cloud based on a memory-centric and heterogeneous multiprocessing architecture, integrated into a high-performance computing data center grade environment. We demonstrate that utilizing the quantum cloud, our hybrid quantum algorithms including Quantum Encoding (QuEnc), Hybrid Quantum Neural Networks and Tensor Networks enable advantages in optimization, machine learning, and simulation fields. We show the advantage of hybrid algorithms compared to standard classical algorithms in both the computational speed and quality of the solution. The achieved advance in hybrid quantum hardware and software makes quantum computing useful in practice today.

Motivation & Objective

• Motivate and build a practical hybrid quantum computing platform integrating quantum and classical resources in a memory-centric HPC environment.
• Demonstrate that hybrid quantum algorithms yield speed and quality advantages in optimization (MaxCut), machine learning (classification and regression), and quantum-inspired simulations.
• Showcase how a unified memory-centric architecture enables efficient quantum-classical interactions and scalability for industrial and research use.

Proposed method

• Introduce a memory-centric HQC cloud (QMware) with in-memory compute and a unified intermediate representation for quantum circuits.
• Use simulated QPUs and native QPUs across topologies via a common SDK to run hybrid quantum algorithms (QuEnc, HQNN, tensor-network approaches).
• Apply QuEnc to MaxCut as a quantum-encoded, amplitude-encoded variational approach and compare with CPLEX under different hardware setups.
• Develop hybrid quantum neural networks for classification and regression benchmarks showing accuracy and data-efficiency gains over classical nets.
• Demonstrate tensor-network-based simulation as a quantum-inspired method with favorable scaling for PDE-like problems.

Experimental results

Research questions

• RQ1 Can a memory-centric hybrid quantum cloud outperform classical solvers on discrete optimization (MaxCut) in practice?
• RQ2 Do hybrid quantum neural networks provide tangible accuracy and data-efficiency gains over classical networks in classification and regression tasks?
• RQ3 Can quantum-inspired tensor networks offer scalable advantages for high-dimensional simulations compared to traditional solvers?
• RQ4 How does integration of simulated QPUs, CPUs, and GPUs within a unified memory model impact end-to-end performance of hybrid quantum applications?

Key findings

• QuEnc with amplitude encoding achieves more accurate MaxCut solutions than CPLEX on 256-node graphs within 1 minute in simulation (costs cited).
• Hybrid pipeline with QuEnc then CPLEX yields improved solutions on 1024-node graphs, surpassing pure CPLEX performance in the reported experiment.
• HQNNs achieve higher test accuracy than classical networks (0.940 vs 0.831) and converge faster (32 vs 317 epochs) on a circle-within-circle classification task.
• Hybrid models for Boston housing regression show 12-16% lower test loss than classical models, indicating robustness with smaller datasets.
• Tensor-network based simulations of Poisson-like problems show exponential or superior scaling advantages over conjugate gradient methods, enabling very large problem spaces (up to 10^9 points) on limited hardware.