An important area of research in quantum computing concerns the application of quantum computers to training of quantum neural networks. The Google AI Quantum team recently published two papers that contribute to the exploration of the relationship between quantum computers and machine learning.
In the first of the two papers, "Classification with Quantum Neural Networks on Near Term Processors", Google researchers propose a model of neural networks that fits the limitation of current quantum processors, specifically the high levels of quantum noise and the key role of error correction.
The second paper, "Barren Plateaus in Quantum Neural Network Training Landscapes", explores some peculiarities of quantum geometry that seem to prevent a major issue with classical neural networks, known as the problem of vanishing or exploding gradients.
InfoQ took the chance to speak with Google senior research scientist Jarrod McClean to better understand the importance of these results and help frame them in a larger context.
InfoQ: Could you elaborate on the importance of the results presented in the two recent papers from the Google AI Quantum team?
Jarrod McClean: The post focuses on two fairly distinct results on quantum neural networks. The first shows a general method by which one may use quantum neural networks to attack traditional classification tasks. We feel this may be important as a framework to explore the power of quantum devices applied to traditional machine learning tasks and problems. In cases where we are only able to conjecture the advantages of quantum computing, it will be important to evaluate the performance empirically and this provides one such framework.
The second result is about the existence of a fundamental and interesting phenomenon in the training of quantum neural networks. It reflects the fact that sufficient randomization in a quantum circuit can act almost like a black hole, making it very difficult to get information back out. However, these traps can be avoided with clever strategies, and the importance of this work was showing when to expect these traps and how to detect them. We believe this knowledge will prove critical in designing effective training strategies for quantum computers.
InfoQ: Can you provide some insights about the directions the Google AI Quantum team is currently heading to with their research? What are your next goals?
McClean: A primary goal of the group is to demonstrate a beyond-classical task on our quantum chip, also known as "quantum supremacy". In parallel, we are interested in the development of near-term applications to run on so-called noisy intermediate scale quantum (NISQ) quantum devices. Development of such algorithms and applications remains a key goal for us, and our three primary application areas of interest are quantum simulation of physical systems, combinatorial optimization, and quantum machine learning.
InfiQ: What is the promise behind machine quantum learning?
McClean: Quantum machine learning has traditionally touted at least two potential advantages. The first is in accelerating or improving the training of existing classical networks. This angle we did not explore in our recent work.
The second is the idea that quantum networks can more concisely represent interesting probability distributions than their classical counterparts. Such an idea is based on the knowledge that there are quantum probability distributions we know to be difficult to sample from classically.
If this distinction between distributions turned out to be true even for classical data, it could mean more accurate machine learning classifiers for less cost or potentially better generalization errors upon training. However I want to be clear that the benefits in this area at the moment are largely conjectured in the case of classical data, and as in the case of classical machine learning, the proof may have to first be empirical. For that reason, in these two works we attempted to enhance the capabilities for testing this idea in practice, in the hope that quantum devices in the near future will allow us to explore this possibility.
Quantum supremacy is the conjecture that quantum computers have the ability to solve problems that classical computers cannot. This is one of the hottest area of research in quantum computing that involve practically all major companies, including Google, IBM, Microsoft, and others. In the case of Google, the company aims to demonstrate quantum supremacy by building a 50-qubit quantum processor and using the simulation of coin tosses as battleground for the proof.
