Abstract
Generative adversarial networks are an emerging technique with wide applications in machine learning, which have achieved dramatic success in a number of challenging tasks including image and video generation. When equipped with quantum processors, their quantum counterparts—called quantum generative adversarial networks (QGANs)—may even exhibit exponential advantages in certain machine learning applications. Here, we report an experimental implementation of a QGAN using a programmable superconducting processor, in which both the generator and the discriminator are parameterized via layers of single- and multi-qubit quantum gates. The programmed QGAN runs automatically several rounds of adversarial learning with quantum gradients to achieve a Nash equilibrium point, where the generator can replicate data samples that mimic the ones from the training set. Our implementation is promising to scale up to noisy intermediate-scale quantum devices, thus paving the way for experimental explorations of quantum advantages in practical applications with near-term quantum technologies.
Introduction
Generative Adversarial Networks (GANs) have revolutionized modern machine learning by enabling the generation of high-quality synthetic data. Quantum Generative Adversarial Networks (QGANs) represent a promising extension of this framework, theoretically capable of providing exponential speedups in learning complex data distributions. However, previous experimental efforts were largely restricted to single-qubit systems and lacked efficient training protocols. This study addresses these limitations by implementing a multi-qubit QGAN on a programmable superconducting processor, utilizing both entanglement and a hardware-efficient quantum gradient descent method.
Experimental Implementation
The experimental platform consists of a five-qubit superconducting processor where qubits are interconnected via a central bus resonator. The QGAN architecture employs two competing Parameterized Quantum Circuits (PQCs): the Generator ($G$), which attempts to produce data mimicking a target source, and the Discriminator ($D$), which seeks to distinguish between real and generated samples.
To optimize performance on near-term hardware, the researchers utilized a hardware-efficient ansatz. Instead of decomposing circuits into standard logic gates, the circuits leverage natural multi-qubit entangling interactions ($U_{ENT}$), resulting in shallower circuits that are more resilient to decoherence. Furthermore, the training process incorporates an on-chip quantum gradient subroutine based on the Hadamard test. This method allows the processor to calculate gradients directly using an auxiliary qubit, offering a significant efficiency advantage over classical finite-difference estimation methods.
Results: Quantum and Classical Data Learning
The QGAN's performance was evaluated on both quantum and classical datasets. In the task of learning an arbitrary mixed quantum state, the system demonstrated exceptional accuracy.
As illustrated in Figure 2, the training process successfully converged to a Nash equilibrium, where the loss function $V$ approached zero. The generator replicated the target mixed state with a fidelity of 0.999.
In a second experiment, the QGAN was tasked with learning the classical truth table of an XOR gate. This required the system to learn discrete classical statistics through adversarial training. After approximately 190 iterations, the system achieved an average state fidelity of 0.927, with the generator producing a truth table nearly identical to the ideal XOR operation. These results confirm the QGAN's ability to handle both continuous quantum states and discrete classical data.
Discussion and Conclusion
The successful implementation of a multi-qubit QGAN highlights the potential of superconducting processors for complex machine learning tasks. The system's stability was verified over 40 hours of continuous operation, proving that current hardware can support the iterative demands of adversarial learning. While scaling to larger systems remains a challenge due to noise in the Noisy Intermediate-Scale Quantum (NISQ) era, the integration of entanglement and efficient quantum gradients provides a robust framework for future explorations of quantum advantage in generative modeling.
Summary updated