By Dr. Chris Mansell
Shown below are summaries of a few interesting research papers in quantum computing and communications that have been published over the past month.
Title: Synthetic weather radar using hybrid quantum-classical machine learning
Organization: Rigetti Computing
Atmospheric conditions are measured all around the world but in oceanic and remote areas, weather radar is not available. In these cases, convolutional neural networks can be trained to produce synthetic radar images, which can then be used to predict the rainfall or the chance of lightning at each pixel on the map.
In a new paper from Rigetti, the researchers investigate whether a hybrid quantum-classical machine learning model can perform well on this operationally relevant task. They started by using theory from a paper called “Power of data in quantum machine learning” and found that the unlabelled weather data was sufficiently complex that a quantum classifier could have an advantage over a purely classical classifier.
The next step was to find a way to actually achieve this advantage for any of the different ways the data could be labelled, e.g. according to the amount of rain or the likelihood of lightning. The theory suggested that their ways of encoding the data into the quantum states wasn’t going to achieve this goal for the available labels.
There is a motto in science: in theory, theory and practice are the same but in practice, they aren’t. With this in mind, they trained their hybrid machine learning system and tested its performance anyway. A quantum variational autoencoder (QVAE) was simulated on a classical computer to start the training, which was then continued on actual quantum hardware, the Rigetti Aspen-9 processor. As an alternative approach, a quantum convolutional neural network (QCNN) was trained on the same quantum device. Both methods employed 32 physical qubits.
The results were that the uncalibrated classical model performed best according to the mean squared error of the predictions. The uncalibrated QVAE trained on satellite data had the highest probability of correctly detecting inclement weather. The QCNN with data about lightning strikes encoded into the rotation angles of its quantum gates outperformed the other approaches on two metrics. Overall, this shows that putting a quantum protocol in a real-world machine learning pipeline can improve its reliability.
Title: Quantum generative adversarial networks with multiple superconducting qubits
Organizations: Nankai University; Chinese Academy of Sciences; Zhejiang University; Tsinghua University; Shanghai Qi Zhi Institute
A generative adversarial network (GAN) is a neural network architecture that when given samples from a specific data distribution, can produce new samples that look as though they came from the input distribution. A concrete example is the generation of photorealistic images of human faces. In this paper, five superconducting qubits with all-to-all connectivity act as a quantum GAN. In one task, it learns to output a specific quantum state with a fidelity of 0.999 and in another, it outputs the state that would result from the application of an XOR gate with a fidelity of 0.927. It achieves these high fidelities by employing multiqubit entangling gates and determining the gradient of its loss function during the training stage. It thus incorporates all of the features needed for a larger scale implementation of this protocol.
Title: Quantum advantage in learning from experiments
Organizations: Caltech; Google; Harvard; Black Hole Initiative; Berkeley; Microsoft; Johannes Kepler University; AWS
During a quantum sensing experiment or a quantum simulation, a quantum computer can learn directly from the quantum data while a classical computer can only process the measurement results. A paper on this topic provides proofs of three different exponential quantum advantages and two proof-of-principle experiments performed on the Sycamore quantum processor with an equal number of system qubits and memory qubits. The first mathematical proof about predicting the values of non-commuting variables was backed up a quantum experiment with an orders of magnitude improvement over a classical recurrent neural network. The second proof was related to quantum principal component analysis and even though there wasn’t a related experiment, it is expected to be robust to noise. The final proof and experiment showed that quantum experiments are significantly better at learning quantum dynamics than classical approaches.
Title: A quantum processor based on coherent transport of entangled atom arrays
Organizations: Harvard; QuEra Computing Inc.; University of Innsbruck; Austrian Academy of Sciences; Massachusetts Institute of Technology; AWS
It is beneficial for the qubits in a quantum computer to have a high connectivity – ideally, all-to-all connectivity – but this is difficult to achieve. Attempts to use a data bus or to move the qubits around have been limited in scale. This paper describes important upgrades to a cold atom experiment that enable coherent transport of the atoms. The researchers showed that they can transport an atom past 2000 other atoms in less than one thousandth of the dephasing time. In combination with newly improved Rydberg-based logic gates, they created various states: a 7-qubit Steane code state; a surface code state with 19 qubits; and a toric code state on a torus with 24 qubits. They also investigated how entanglement changed with time in a non-trivial way.
Title: Randomized Compiling for Scalable Quantum Computing on a Noisy Superconducting Quantum Processor
Organizations: University of California at Berkeley; Lawrence Berkeley National Lab; Massachusetts Institute of Technology; Quantum Benchmark Inc.; Keysight Technologies Canada; University of Waterloo
Systematic mistakes in the control of qubits cause unitary errors that can interfere with each other in ways that are hard to predict. This means that the worst-case scenario, where the errors constructively interfere, can be orders of magnitude worse than the average situation where there would be some partial cancelling of the errors. Fortunately, in 2016, theorists devised a technique called randomised compiling to deal with these coherent errors. In this month’s paper, experimentalists working with four superconducting qubits have demonstrated that the protocol offers significant performance gains. Furthermore, they showed that as quantum computing hardware improves and gate fidelities increase, the protocol will be even more effective.
Title: Revisiting dequantization and quantum advantage in learning tasks
Organizations: Caltech; Google; Harvard; Black Hole Initiative
Quantum computing is such a complex topic that despite all the research that has gone into it, some very foundational issues are still open. In particular, the comparison between quantum and classical algorithms is only just starting to be clarified now. Take principal component analysis as an example. The quantum algorithm that gets quantum states as its input has an exponential advantage over a classical algorithm with access to measurement data but this is not the case if the classical algorithm has sample and query access to the vectors describing the quantum state. The authors of this paper have brought some much-needed clarity to these comparisons by concentrating on how the data is accessed. This helpful work should allow discussions of quantum advantages to be much more straightforward.
Title: Differentiable quantum computational chemistry with PennyLane
Organizations: Xanadu; University of Toronto; AWS; Vector Institute for Artificial Intelligence; Harvard; Canadian Institute for Advanced Research
Specialised software libraries make it possible for programmers to use advanced techniques without having to look through all the nitty gritty details. Recently, software tools for automatic differentiation have helped reduce the mathematical prerequisites for machine learning. This paper describes how the Python library called PennyLane offers similar functionality to coders of quantum computers, particularly those interested in the interface between quantum machine learning and quantum chemistry. Quantum circuits for computing the ground state energy of a molecule are often designed with variable parameters. These circuits can be differentiated with respect to their free variables just like neural networks can be with respect to their weights. By providing increasingly powerful features with a simple user interface, PennyLane is contributing to the advancement of quantum chemistry.
Title: Quantum Variational Optimization of Ramsey Interferometry and Atomic Clocks
Organizations: University of Innsbruck; Institute for Quantum Optics and Quantum Information of the Austrian Academy of Sciences; Leibniz University Hannover
Constraints limit success and when it comes to metrology, the limit on the achievable signal-to-noise ratio is higher when a system can make use of entanglement. The difficulty of a task is also a factor: in situations where the value being measured is stable, GHZ states are known to be optimal but the authors of this paper consider the harder challenge of determining a fluctuating variable with a single-shot measurement. They classically simulated and optimised some low-depth variational quantum circuits and found performance approaching the fundamental limits. A preprint from a few months ago implemented this work on 26 trapped ions and demonstrated 2 dB of metrological gain over classical interferometers. Ultimately, the idea to combine a variable quantum circuit with a precision measurement protocol should also be applicable and beneficial to state of the art atomic clocks.
Related hardware preprint link: https://arxiv.org/abs/2107.01860
Title: Measuring the Capabilities of Quantum Computers
Organizations: Sandia National Laboratories
Benchmarking allows potential users of quantum computers to know whether a particular device will be capable of performing their desired application. For example, IBM’s quantum volume metric makes it clear if a particular unitary can be applied and Google used cross-entropy measures to make conclusions about the difficulty of sampling from a specific distribution. When it comes to quantum circuits, the main approach to benchmarking is to try random circuits but, as pointed out in the paper, this approach has poor scaling and is insensitive to structured, coherent noise. The authors of the paper have created a scalable way to investigate the interplay between the structure in a chosen protocol’s circuit and the structure in the noise that a given quantum processor experiences. They used their new benchmarking methods to reach interesting conclusions about Rigetti and IBM machines.
December 27, 2021