Artificial intelligence running on a quantum computer can now generate recognisable images of things like shoes and T-shirts, using the same methods as popular text-to-image tools like Dall-E or Midjourney. They still aren鈥檛 what you would call stunning images, but if the method scales up to more powerful machines, it should lead to much higher-resolution pictures.
Generative adversarial networks (GANs) are a kind of AI that pits two neural networks against each other. A network called a generator creates artificial data, such as an image, that is as close to the real data as possible; a second network, the discriminator, must work out whether the image is real or generated.
As this process is repeated many times, the generator learns to produce better-quality images that can fool the discriminator into thinking they are real.
Advertisement
There have been attempts to implement this process on quantum computers, which promise to provide exponentially more computing power, but they have resulted in very poor images and the methods have failed to scale up to more powerful quantum machines.
Now, at Northeastern University in Massachusetts and his colleagues have developed a quantum GAN called MosaiQ, which can reliably generate recognisable images and should scale up as more powerful quantum machines become available. 鈥淓ven with such limited resources, we show how well this performs,鈥� says Silver.
In MosaiQ, the generator is quantum while the discriminator is classical. This still offers a significant advantage over classical GANs, says Silver, because of the way that images are represented on computers and how they can be mapped to quantum circuits, which have the benefit of storing data in many more states, or dimensions, than the 1s and 0s of classical machines.
鈥淎ll the pixels are related to each other, in terms of their positioning and what colour they have, or in this case what sort of amplitude or intensity they have,鈥� says team member at Rice University in Texas. 鈥淭hese pixel relationships can be better emulated using these higher-dimensional quantum networks.鈥� This means quantum generators should be able to produce higher-quality images.
The researchers used IBM鈥檚 7-qubit Jakarta quantum computer to run MosaiQ for two different data sets of black and white images, one of the digits 1 to 9 and another of basic clothing items like shoes, T-shirts and trousers. Unlike in previous work, the artificial images generated by MosaiQ are hard to distinguish from the original data set.
While the image quality is currently just 28 x 28 pixels (see photo, top), Patel says the software should be easily usable on more powerful quantum machines.
arXiv
Article amended on 9 October 2023
We have corrected the name of the lead author of this research.
