Efficient Machine Learning Systems for High-Fidelity Qubit Readout

Multi-qubit readout is among the most error-prone operations in superconducting quantum computing systems. These errors occur for various reasons, including but not limited to: crosstalk between the readout tones in a frequency-multiplexed readout scheme, spontaneous state transitions during the measurement, excitations caused by the readout pulse, and thermal noise added to the readout signal as it travels from the refrigerator to the room-temperature electronics. Prior works on reducing readout errors include machine learning-assisted readout, where a neural network is used for more robust discrimination by compensating for crosstalk errors. However, the neural network size can limit systems' scalability, especially if fast hardware discrimination is required. This work presents a scalable approach for mitigating single-shot readout errors by using a matched filter in conjunction with a significantly smaller and scalable neural network for qubit-state discrimination. In addition, we optimize the training of the matched filter and neural network by using a pre-classifier that filters incorrectly labeled training data. Fast and accurate discrimination of qubit states is essential for deploying quantum error correction codes. To that end, we investigate computationally efficient and scalable machine learning algorithms for enabling high-fidelity multi-quit readout that are readily implementable on FPGAs.