Neural Network assisted Superconducting Qubit Readout*

A significant error source in contemporary quantum processors is qubit-state readout. For a single qubit connected to a unique readout line, a linear matched filter is sufficient for high-fidelity readout. However, it is more resource efficient to frequency-multiplex multiple qubits on the same readout line. In this case, the single-qubit matched filters are no longer optimal. Rather, readout discrimination becomes a computationally intensive, multi-state classification problem. Here, we present a new approach to the readout problem based on neural networks. We discuss different types of neural network architectures and their readout discrimination performance compared to current readout methods when applied to experimental single and multi-qubit readout data.