Quantum Generative Adversarial Networks QGAN - RPIQuantumComputing/QuantumCircuits GitHub Wiki

Quantum generative adversarial networks or QGAN is an extension of classical generative adversarial networks and is a powerful machine learning algorithm that is used mainly for the generation of synthetic data into quantum mechanics and concepts, and as usual, the extension into quantum mechanics in theory is able to enhance the capabilities of the two main parts of a generative adversarial networks in order to be able to compute larger data points.

First of all, lets explain what a classical generative adversarial networks is, a classical GAN is made up mainly of two components: a generator and a discriminator, these two components are basically two neural networks that are engage in a competitive process. The generator is mainly used for creates data instances that tries to mimic actual data, while the discriminator tries to discriminate between real data and the data generated by the generator. This relationship is what compels the generator to create increasing accurate data points that makes it harder and harder as time goes by for the discriminator to determine which data is real and which data is false. Thus as the training continues, the generator becomes more accurate and it would evenly become adept at creating synthetic data and this process is used mainly for image generation, style transfer, text to image synthesis, and more. The GAN isn't flawless, in fact, it has many faults, some of the most glaring ones include issues in the training, where if no equilibriums are reached between the generator and the discriminator, the generator might suffer from mode collapse which basically means the generator does not go through the entire input data sets and the synthetic data generated are all very visibly false. Another issue GAN has is how expensive it is to train, the training of a GAN may require expansive hardware and extremely long training times and training iterations to achieve results of satisfactory level. These two are only some errors in the training and implementation of GAN.

A quantum generative adversarial network, unlike its classical counterpart, is able to sidestep some of the issues presented in the classical generative adversarial networks by taking advantage of quantum concepts and mechanics, solve problems that classical GANs cannot solve. There are many differences between classical and quantum GANs, but the first difference between the two is that the two components, the generator and the discriminator are that quantum circuits instead uses parametrized quantum circuits instead of neural networks, though it possible to build a quantum GAN using a quantum neural network or anything with a complete parametrized rotational gate, this design causes the input state to go through a unitary operation and creates a quantum state that depends on the parameter of the circuit. The discriminator of the quantum GAN, evaluates both classical training data samples as well as quantum data samples generated by the generator circuit, and because of this difference, the discriminator has to match the basis state to the training data's feature space which results in a classical synthetic data point that can be compared to the classical training data, another method of comparison is used by the generator to convert classical data into quantum data using a superposition over the different basis states and is given by an amplitude times the basis states. The objective of the QGAN remains the same when compared to its classical counterpart, the discriminator still aims to differentiate between real and fake training data while the generator still aims to create good data points that closely resembles the actual data.

When comparing the quality of quantum GAN with classical GAN, quantum GAN trumps its classical counterpart when faced with a large amount of data points, the quantum GAN be much faster during the training process due to taking advantage of quantum mechanics of entanglement and superposition to enhance the generation and the discrimination of the quantum samples and quantum effects can also be leveraged to generate more diverse and realistic quantum data. Despite the advantages that quantum GAN has, there are still some errors that held it back, some of these problems include that quantum GAN is susceptible to noise and outliers or a lack of training data, where a GAN can be trained on readily available large scale datasets, quantum GAN does not have that luxury, and many more problems that prevents it from being mainstream.