Quantum Generative Adversarial Networks (QGAN)
Quantum Generative Adversarial Networks (QGAN)
A quantum generative adversarial network (QGAN) is a game between a generator and a discriminator, one or both of which are variational quantum circuitsVariational Quantum Circuits: A variational quantum algorithm optimises the parameters of a quantum circuit U(θ) in a classical loop to minimise a cost function estimated from measurements, such as t…. The generator learns to produce states or samples that the discriminator cannot distinguish from the data.
An early research model: presented for learning and research purposes, with no general evidence of its superiority in practical applications.
Lloyd & Weedbrook (2018) formulated the idea theoretically, and Dallaire-Demers & Killoran (2018) presented a design with variational circuits. Its main forms:
- Quantum generator and classical discriminator: the generator is a circuit that is measured, giving discrete samples judged by a classical discriminator. Zoufal and colleagues (2019) used this form to learn probability distributions and load them into the circuit.
- Quantum generator and discriminator: both operate directly on quantum states, suited to learning quantum data.
- Hybrid GANs for image generation: such as patch GAN generators that produce very small images, e.g. 8×8, with several small circuits.
Training alternates between updating the discriminator to improve its discrimination and updating the generator to fool it. It shows the same difficulties as classical GANs, such as instability and mode collapse, plus shot noise and barren plateausBarren Plateaus: Barren plateaus are a phenomenon in which the variance of the cost-function gradient vanishes exponentially with the number of qubits in deep random parameterised circui….
Realistic assessment: quantum-generated images today are very small and simple, and do not compare with modern classical generative models.
Mathematical formulation
Worked example
Exercise: design a two-qubit generator with RY rotations and CNOT, and a target distribution such as p(00) = 0.5 and p(11) = 0.5. Which angles achieve this? (Answer: RY(π/2) on q0 then CNOT, which is a Bell state.) In a QGAN the generator finds such angles through adversarial training.
Try it yourself: Bell State Experiment
Test your understanding
The quiz needs JavaScript.
References
- Seth Lloyd; Christian Weedbrook (2018). Quantum Generative Adversarial Learning. Physical Review Letters 121, 040502 doi:10.1103/PhysRevLett.121.040502
- Pierre-Luc Dallaire-Demers; Nathan Killoran (2018). Quantum generative adversarial networks. Physical Review A 98, 012324 doi:10.1103/PhysRevA.98.012324
- Christa Zoufal; Aurélien Lucchi; Stefan Woerner (2019). Quantum Generative Adversarial Networks for learning and loading random distributions. npj Quantum Information 5, 103 doi:10.1038/s41534-019-0223-2
