Deep-Neural-Network Discrimination of Multiplexed Superconducting-Qubit States

Abstract

Demonstrating a quantum computational advantage will require high-fidelity control and readout of multiqubit systems. As system size increases, multiplexed qubit readout becomes a practical necessity to limit the growth of resource overhead. Many contemporary qubit-state discriminators presume single-qubit operating conditions or require considerable computational effort, limiting their potential extensibility. Here, we present multiqubit readout using neural networks as state discriminators. We compare our approach to contemporary methods employed on a quantum device with five superconducting qubits and frequency-multiplexed readout. We find that fully connected feedforward neural networks increase the qubit-state-assignment fidelity for our system. Relative to contemporary discriminators, the assignment error rate is reduced by up to 25% due to the compensation of system-dependent nonidealities such as readout crosstalk, which is reduced by up to one order of magnitude. Our work demonstrates a potentially extensible building block for high-fidelity readout relevant to both near-term devices and future fault-tolerant systems.

Transmon qubit readout fidelity at the threshold for quantum error correction without a quantum-limited amplifier

“…To further develop these readout techniques, more complex methods in the construction of the primary and secondary readout tones could be explored. More sophisticated deep neural-network methods could also be employed to aid state classification of the two-tone readout results [49]. The possibility to generalize these techniques to further boost fidelity for multiplexed readout is a promising prospect for the future investigation of quantum computing with superconducting qubits.…”

“…In the design of future devices, the qubits could be grouped into physically separated readout lines depending on their designation, e.g., ancilla or data qubits, and whether their measurements occur simultaneously. Moreover, the induced crosstalk could be further mitigated with other techniques such as machine-learning algorithms for discrimination and readout pulse shaping [49,51,52].…”

“…The second method to discriminate the qubit state utilizes a feedforward neural network (FNN). The network was organically developed for multiplexed readout [49] and is adapted here to treat the two data sets as a single system (see Methods section). The input to the neural network combines the in-phase (I[n]) and quadrature (Q[n]) data from the n th integrated signal of the two tones into a four-element vector…”

“…We utilize a feedforward neural network (FNN) with two hidden layers to discriminate the qubit state using the combined two-tone results. The choice of using an FNN over other methods such as support-vector machine (SVM) or non-linear support-vector machine (NSVM) is justified by the fact that an FNN could achieve greater performance when discriminating more than two states as well as better scalability [49]. An SVM or NSVM will also need to be retrained from scratch every time.…”

“…While on the other hand, the FNN is capable of transfer learning, where retraining of the network during future re-calibration of the system is significantly more efficient [55]. The advantage of using FNN is significant enough to warrant a wider application [49,51,56].…”