Neural networks for multiqubit readout

Short answer: Reading several superconducting qubits through one shared line produces crosstalk between channels, and a neural network can be trained to tell the qubit states apart despite it. A 2022 Physical Review Applied paper, with an arXiv preprint version, reports that feedforward neural networks cut the assignment error rate by up to 25% on a five-qubit device with frequency-multiplexed readout, and cut readout crosstalk by up to one order of magnitude. This is one experiment on one device.

The readout problem

The abstract says demonstrating quantum computational advantage will require high-fidelity control and readout of multiqubit systems. As systems grow, multiplexed readout, where several qubits share measurement hardware, becomes a practical necessity to limit growth in resource overhead. The authors say many contemporary qubit-state discriminators presume single-qubit operating conditions or need considerable computational effort, which limits how far they extend.

What the authors did

The authors present multiqubit readout using neural networks as state discriminators. They compare it with contemporary methods on a quantum device with five superconducting qubits and frequency-multiplexed readout. A discriminator takes the measured readout signal and decides which state each qubit was in, and the fraction it gets right is the assignment fidelity.

What they found

They find that fully connected feedforward neural networks increase the qubit-state-assignment fidelity for their system. Relative to contemporary discriminators, the assignment error rate is reduced by up to 25%. The authors attribute this to the compensation of system-dependent nonidealities such as readout crosstalk, which they report is reduced by up to one order of magnitude. They describe the work as demonstrating a potentially extensible building block for high-fidelity readout relevant to both near-term devices and future fault-tolerant systems. The extensibility is their expectation and was not tested beyond five qubits in the abstract.

Why a trained discriminator can help

A neural network can learn the particular distortions of one device, including how a signal from one qubit leaks into another's channel. The authors say this is the reason for the gain, since the network compensates for nonidealities specific to the system. The flip side is that the network is trained on that system, so the result describes that device and its calibration.

What the numbers mean

An assignment error rate is the fraction of readouts where the discriminator names the wrong state. A 25% reduction in that rate means, for example, that an error of 4 in 100 readouts would become 3 in 100, though the abstract does not give the underlying rates, so no absolute figure should be assumed. Crosstalk is a separate quantity, how much one qubit's readout signal contaminates another's, and its reported tenfold reduction at the best case is a statement about the signals, not directly about gates.

Limits

The reported gains are upper bounds, "up to 25%" and "up to one order of magnitude", for a five-qubit device. The abstract does not state results on larger devices or on how often retraining is needed. The arXiv version is the same work as the journal paper, so the two are not independent confirmation.

FAQ

What is multiplexed readout?

Measuring several qubits through shared hardware, such as one line carrying signals at different frequencies.

What did the neural networks improve?

The qubit-state assignment fidelity, with the assignment error reduced by up to 25%.

What happened to crosstalk?

Readout crosstalk was reduced by up to one order of magnitude, per the authors.

How big was the device?

Five superconducting qubits.

Does it scale?

The authors call it potentially extensible. The abstract does not test larger systems.