Abstract
Machine learning techniques have led to broad adoption of a statistical model of computing. The statistical distributions natively available on quantum processors are a superset of those available classically. Harnessing this attribute has the potential to accelerate or otherwise improve machine learning relative to purely classical performance. A key challenge toward that goal is learning to hybridize classical computing resources and traditional learning techniques with the emerging capabilities of general purpose quantum processors. Here, we demonstrate such hybridization by training a 19-qubit gate model processor to solve a clustering problem, a foundational challenge in unsupervised learning. We use the quantum approximate optimization algorithm in conjunction with a gradient-free Bayesian optimization to train the quantum machine. This quantum/classical hybrid algorithm shows robustness to realistic noise, and we find evidence that classical optimization can be used to train around both coherent and incoherent imperfections.
Introduction
In the current "Noisy Intermediate-Scale Quantum" (NISQ) era, quantum hardware is characterized by significant error rates and limited coherence times. These constraints make the implementation of deep, error-corrected algorithms like Shor’s algorithm currently impractical. However, near-term quantum devices can still provide value through hybrid architectures. This paper explores the application of a 19-qubit quantum processor to unsupervised machine learning, specifically focusing on the clustering problem. By treating clustering as a combinatorial optimization task, the researchers demonstrate how quantum and classical resources can be effectively integrated to solve problems that are difficult for classical machines alone.
Hybrid Quantum-Classical Architecture
The core of this work is a hybrid loop that partitions the computational workload between a quantum processor and a classical optimizer.
The process involves two primary stages:
- Quantum Stage (QAOA): The clustering problem is mapped to a Maximum-Cut (Maxcut) problem and encoded into an Ising Hamiltonian. The Quantum Approximate Optimization Algorithm (QAOA) is then executed on the Rigetti 19Q processor. This involves preparing a quantum state through a parameterized circuit defined by angles $\gamma$ and $\beta$. Measuring this state yields bit-strings representing potential cluster assignments.
- Classical Stage (Bayesian Optimization): Because quantum measurements are inherently noisy, standard gradient-based optimization methods often fail. Instead, a gradient-free Bayesian optimizer is employed. It uses the quantum output to update a Gaussian Process model and selects the next optimal parameters ($\gamma, \beta$) to minimize the cost function.
The hardware utilized was the Rigetti 19Q, a superconducting gate-model processor. To maximize performance, the problem instances were mapped to the chip's native lattice connectivity, reducing the need for additional noisy "swap" gates.
Experimental Setup and Datasets
The study evaluated the hybrid algorithm using synthetic datasets:
- 19-point Clustering: Points were selected to align with the 19Q chip's physical connectivity.
- 20-point Fully Connected Graph: Simulated on the Rigetti Quantum Virtual Machine (QVM) to assess performance without hardware connectivity constraints.
- Similarity Metric: The Bhattacharyya Coefficient was used to construct distance matrices, which were then converted into the cost Hamiltonian.
Results and Performance
The hybrid approach demonstrated significant robustness and efficiency. A key finding was the ability of the classical optimizer to "train around" hardware noise. By treating the 19Q chip as a black box, the Bayesian optimizer identified parameters that compensated for both systematic (coherent) and random (incoherent) errors.
Statistical analysis using the Kolmogorov-Smirnov test confirmed that the hybrid algorithm outperformed random bit-string sampling with over 99% confidence. In simulations of a 20-qubit system, the algorithm achieved a 100x improvement in sampling efficiency compared to random guessing. Most experimental runs on the physical hardware converged to high-quality approximate solutions well within the allotted optimization steps.
Discussion and Future Outlook
While successful, the study highlights several challenges for scaling. The experiments were restricted to a single QAOA iteration ($p=1$); increasing the depth would theoretically improve solution quality but also amplify the impact of noise. Furthermore, as the number of parameters grows with circuit depth, the classical optimizer's computational cost—specifically the $O(m^3)$ scaling of Gaussian Process updates—may become a bottleneck.
Future improvements in gate fidelity and qubit connectivity will be essential. However, this demonstration proves that hybrid quantum-classical systems can already solve non-trivial machine learning problems on noisy hardware, paving the way for more complex applications in high-dimensional data analysis.
Summary updated