By Dr. Chris Mansell

Hardware

Title: Two-qubit silicon quantum processor with operation fidelity exceeding 99%
Organizations: Princeton; NIST / University of Maryland; Sandia National Laboratories

In January of this year, some very high logic gate fidelities for silicon qubits were reported in Nature. However, quantum computers also need to have high state preparation and measurement (SPAM) fidelities. In silicon devices, SPAM fidelities have ranged from about 80 to 90%, so the imperative was to improve this value. In April, new experimental work has pushed it to above 97%. The logic gate fidelities, in ascending order, were similarly impressive: 99% for single-qubit gates operating in parallel, 99.8% for a two-qubit controlled-phase gate and 99.9% when controlling just one qubit at a time. Given these numbers and some recent advances in fabrication methods, the paper, published in Science, claims that silicon spin qubits could scale up to become a “dominant” technology for NISQ processors. 
Link: https://www.science.org/doi/10.1126/sciadv.abn5130

Title: Experimental photonic quantum memristor
Organizations: University of Vienna; Politecnico di Milano; Consiglio Nazionale delle Ricerche

A memristor is a nanoscale electronic component with the peculiar property that its resistance depends on how much charge has passed through it. Classical memristors are considered extremely exciting candidates for next generation computing systems. The quantum memristor proposed and demonstrated in this article obeys the same equations as its classical counterpart but it can coherently operate on quantum states. In the device, photons travel down different paths of an integrated photonic chip. When they are measured, an active feedback loop is used to tune the reflectivity of the chip’s beamsplitters. The non-linearity and short-term memory that this provided were then put to use in an image classification task. When single-photons encoded the data, the experiment was more accurate and more resource efficient than has been reported for classical memristors. These are encouraging results and it will be interesting to see how future, scaled-up versions of this set-up perform.
Link: https://www.nature.com/articles/s41566-022-00973-5

Title: High coherence and low cross-talk in a tileable 3D integrated superconducting circuit architecture
Organizations: University of Oxford; University of Southampton

This paper reports on a two-dimensional tile containing four superconducting transmon qubits and shows that its design means numerous tiles could be placed into a large grid without degrading any performance metrics. Firstly, the wires are placed on the opposite side of the tile to the qubits and they pop out of the plane so that they can avoid each other. Secondly, the tile is enclosed in a cavity that provides a clean electromagnetic environment for the qubits. The single-qubit gate fidelities were found to be 99.982% regardless of whether the gates were performed sequentially or simultaneously. This implies that the cross-talk is extremely low and helps build the case that the tile is scalable. Several calculations in the supplementary information indicate that when more tiles are added and the qubits get wired together to allow two-qubit gates to be performed, the coherence times and the fidelities should remain constant.
Link: https://www.science.org/doi/10.1126/sciadv.abl6698

Title: Quantum state preparation and tomography of entangled mechanical resonators
Organization: Stanford University

A quantum acoustic processor with coupling between a superconducting transmon qubit and a pair of piezoelectric, nanomechanical, phononic crystal resonators might sound like something from a sci-fi novel. On the contrary, it is an experimentally demonstrated device that has been rapidly improving in recent years. In this research, improved fabrication methods extended the coherence times of the transmon qubit and the resonators. A deterministic iSWAP operation was implemented in about 25 nanoseconds with an estimated fidelity of 95% and quantum non-demolition measurements were performed for the read-out. The plan is to further extend the coherence times before incorporating features like quantum random access memories or Kerr-cat qubits into the architecture.
Link: https://www.nature.com/articles/s41586-022-04500-y

Software

Title: An analytic theory for the dynamics of wide quantum neural networks
Organizations: The University of Chicago; Chicago Quantum Exchange; IBM; University of Maryland

This paper examines a variational quantum algorithm performing supervised learning and considers how its residual training error (RTE) varies as it gets trained by gradient descent. In order to stand a chance at arriving an analytical result, the authors had to consider overparameterised regimes, where the number of trainable parameters is large. They used the theory of neural tangent kernels to find that the RTE could converge exponentially. These results were of immediate practical interest because they shed light on the findings of some previously published numerical observations. Furthermore, the authors were able to identify types of quantum circuits that could have these desirable, exponentially converging, training dynamics without an unmanageably large amount of parameters. Going forwards, they plan to explore different circuit designs and see if there are any links to how well the learning algorithms can generalise when given new data.
Link: https://arxiv.org/abs/2203.16711

Title: ADAPT-VQE is insensitive to rough parameter landscapes and barren plateaus
Organization: Virginia Tech

The variational quantum eigensolver, or VQE, is an algorithm for computing the energy levels of molecules. This paper focusses on an adaptive and problem-tailored version of the protocol known as ADAPT-VQE, where instead of optimising the variational parameters of the circuit, it is the circuit structure itself that is iteratively improved. Numerical simulations of several molecules showed that the initial guesses of ADAPT-VQE help it avoid local minima. Even if the algorithm does find itself in a troublesome parameter regime, such as a local minimum or a barren plateau, it can “burrow” towards the exact solution. Some challenges remained, specifically one which the researchers termed “gradient troughs,” so the immediate plan is to investigate these further.
Link: https://arxiv.org/abs/2204.07179

Title: Out-of-distribution generalization for learning quantum dynamics
Organizations: Technical University of Munich; Munich Center for Quantum Science and Technology; Freie Universit ̈at Berlin; Caltech; Los Alamos National Laboratory; University of Southern California; AWE
Quantum neural networks (QNNs) continue to be a popular research topic with a new preprint providing the first example of a very useful phenomenon. Before starting any machine learning project, one usually accesses a dataset and splits it into two parts, one for training the model and one for testing it. After training, if the model gets a high score on the test data, we say it has shown in-distribution generalisation. Nevertheless, the model can be brittle and get low scores when tested on alternative or modified data. In that case, we say that it failed to generalise out of the distribution. In this latest work, a QNN learns how unitary operations act on separable states and then gets asked how they would operate on entangled states. Both analytical and numerical results showed that the QNN generalised very well from the training to the testing distribution.
Link: https://arxiv.org/abs/2204.10268

Title: Quantum Self-Supervised Learning
Organizations: University of Oxford; University of Cambridge

Machine learning (ML) has been incredibly successful in the past decade, so the current trend in quantum computing is to see what ML ideas could be ported into the quantum domain. Recently, self-supervised learning has become a prominent way to train ML models without the need for human-annotated data. While models that supervise themselves using unlabelled data sound great, the downside is that they are extremely computationally expensive. The idea expressed in this paper is that since quantum neural networks (QNNs) operate in high-dimensional Hilbert spaces, they may be able to learn the complex patterns required for self-supervised learning. The researchers simulated a QNN without any noise or decoherence and showed that it could help an ML model achieve a higher self-supervised accuracy than a purely classical model. When they ran the QNN on IBM’s “Paris” quantum computer, the accuracy equalled that of the classical case, showing that it was robust to the imperfections of a NISQ device.
Link: https://iopscience.iop.org/article/10.1088/2058-9565/ac6825

April 27, 2022