Many scientists believe that a quantum computing device will one day perform highly complex computational tasks which are far beyond the ability of today’s classical computers.
This theory is known as Quantum Supremacy, and scientists are engaged in various research efforts to advance it. A recent publication has created quite a buzz in the quantum community: A Tsinghua University research paper has for the first time reported an experimental demonstration of quantum generative adversarial learning in a superconducting quantum circuit. The trained quantum network, QGAN, achieved an impressive 98.8 percent average accuracy in generating quantum data that is indistinguishable from real data.
The paper Quantum Generative Adversarial Learning In A Superconducting Quantum Circuit was first submitted to ArXiv last August, and has now been published in the respected multidisciplinary open-access scientific journal Science Advances.
What is Quantum GAN?
Generative Adversarial Networks (GANs) are today’s hottest machine learning subfield. In 2014, Google Researcher Ian Goodfellow first introduced the GAN framework, which comprises two neural nets: A Generator that forges a new data instance, and a Discriminator that then distinguishes fake data created by the Generator from real data. In effect, the two neural nets challenge each other with increasingly realistic fakes, both optimizing their strategies until their generated data is indistinguishable from the real data.
While GANs have been widely adopted in image generation and transfer techniques and can now produce highly convincing fake images of animals, landscapes, human faces, etc.; scientists doubt whether GANs run on classical computers will be capable of efficiently solving challenging problems that involve high-dimensional data (with hundreds or thousands of features).
In spring 2018 a pair of researchers theoretically proposed a quantum version of GANs, or Quantum GANs (QuGANs). QuGANs use quantum computing devices to train their generator and discriminator, and can process high-dimensional data, or quantum data. The experiments concluded that:
- Generative models can be trained on quantum computers.
- QuGANs can reach a point where their generator can produce data that is indistinguishable from the real data. Small QuGANs can perform gradient descent to adjust parameters.
- When the data consists of samples of measurements made on high-dimensional spaces, QuGANs may exhibit an exponential advantage over classical adversarial networks.
The researchers tested their QuGAN theory at the time, but its efficacy still required further proof through experiments — which is where the new Tsinghua University paper plays a useful role.
How was the experiment conducted?
The experiment used the same GANs framework, but re-wrapped the data, the generator and the discriminator with quantum mechanics. The Tsinghua researchers’ Quantum GAN (QGAN) includes three components: a black box quantum process for the quantum true data σ, the generator (G) that produces an ensemble of pure quantum states ρ, and the discriminator (D) that performs projective measurements M. Then the two nets engage in an adversarial game until the generator can completely fool the discriminator.
Researchers explain that the QGAN learning algorithm was realized in a “superconducting quantum electrodynamics architecture.” Just one superconducting transmon qubit is involved in the experiment, “dispersively coupled to a bosonic microwave mode.” The optimization of the generator and discriminator was conducted on field programmable gate arrays (FPGA), and the gradient descents performed in a classical computer.
The result demonstrated the feasibility of QGANs with a superconducting quantum circuit: In 98.8 percent of the cases, the discriminator could not distinguish between the real and the generated data.
What are the implications?
Researchers believe the experiment can have far-reaching influence: “Since our QGAN experiment requires neither a quantum random accessing memory, nor a universal quantum computing device or any fine-tuning parameters (thus robust to experimental imperfections), it carries over to the noisy intermediate scale quantum (NISQ) devices widely expected to be available in the near future.”
The term NISQ was coined by John Preskill, the Caltech physicist who also proposed Quantum Supremacy. NISQ technology represents types of quantum computers that have fifty to hundreds of qubits and cannot perform error correction. These computers are expected to become available in the next few years: IBM’s 50-qubit machine and Google’s 72-qubit quantum chip both fall into this category.
Since QGAN was successfully trained on a superconducting quantum circuit — which is considered a NISQ device — the experiment suggests quantum supremacy is achievable on near-future devices with practical applications.
Other implications include:
- Solving quantum many-body problems with QGAN, given the recent rapid progresses in related directions.
- The hybrid quantum-classical architecture demonstrated in this work can be straightforwardly extended to the optimal control and self-guided quantum tomography
- More applications in other quantum machine learning/AI algorithms are expected.
Click here to view the new QGAN paper.
Journalist: Tony Peng | Editor: Michael Sarazen