by Zapata Computing 17 September 2019
Generating data – images or videos – has been an important element of AI research. After all, teaching a machine to come up with new images not only endows the android with a sense of creativity – precious in the quest for Artificial General Intelligence – but also opens the door to many business applications. Generative models are however hard to train and to deploy, but they could be improved through quantum computing.
Quantum computers (QC) are small today and are not likely to replace most AI/ML algorithms in the near future because they simply cannot process the massive amounts of data that are required to train useful models. For instance, QCs will not likely have much of an impact on discriminative models, as the training data is typically a massive set of classically encoded data (though there is much potential in areas like chemistry, where the data is coming from a quantum system).
We are, however, discovering ample opportunities for QCs to augment existing models. Our insight at Zapata Computing is to rely on the quantum computer only for the hardest computational subtask, and allow the classical computer to do the rest. In particular, one opportunity of a quantum computer comes from the fact that they are able to generate probability distributions that classical computers cannot efficiently generate. In fact, within the next year or two, we will likely see an announcement of so-called “quantum supremacy”, where a quantum machine is able to quickly generate complex distributions that no classical computer anywhere could replicate.
One such hybrid system that could be accelerated is a variant of a Generative Adversarial Network (GAN). A GAN is a neural network that works in a way analogous to counterfeiting; one part of the network (the “generator”) acts like a money forger and the other part (the “discriminator”) tries to sort out the fake money from the real money. GANs have been particularly successful in image generation, with applications from astrophysics to fashion modeling. One limitation of GANs is that discriminators tend to improve much faster than generators, and this slows down the training process.
An associate adversarial network (AAN) is a variant of a GAN which tries to improve on the model by introducing a kind of eavesdropper into the mix—more precisely, using a Restricted Boltzmann Machine (RBM) to feed some information from the discriminator to the generator, speeding up its learning. One opportunity, then, is to use a Quantum Boltzmann Machine to replace this classical component. As this part of the neural network has a relatively small number of neurons, it is possible to leverage quantum computing to accelerate this subtask and speedup the overall training time of the AAN.
As quantum computers grow in size, speed, and capability, we envision more opportunities and faster improvements. We feel certain there will come a time when AI is increasingly reliant on quantum computing, taking on a complementary role similar to the role of GPUs and FPGAs today.
The eminent Carl Sagan once said, “We are a way for the Cosmos to know itself.” Quantum computers will be key to taking that knowledge to the next level.