Tackling Quantum Sampling Noise in Quantum Machine Learning on Large-scale Quantum Devices

A new quantum machine learning algorithm, Noise Intermediate-Scale Quantum Reservoir Computing (NISQRC), was recently introduced that takes advantage of features generated by projectively measured quantum systems, where the number of features scales exponentially with the qubit count [1,2]. Designed to handle both time-independent data [1] and streaming time-dependent data [2], the NISQRC algorithm can also be optimized for limited measurement resources (e.g., shot count) through the Eigentask Analysis introduced in Ref. [1]. A recent application of NISQRC on superconducting quantum processors has demonstrated inference on arbitrarily long time-dependent data, unconstrained by coherence time limitations without error mitigation or correction [2]. Its design for particular ML tasks can be systematically optimized using the Quantum Volterra Theory introduced in Ref. [2]. In this study, we numerically investigate the effectiveness of Eigentask Analysis scales with the number of qubits on large multi-qubit devices. Specifically, we identify the count of noise-resilient eigentasks as the number of qubits and measurement shots increases and conduct a comparative analysis using experimental data from an analog quantum computer on a neutral atom platform.