Quantum noise unexpectedly improves neural network performance in experiment on three platforms

Scientists from the Joint Quantum Institute tested a neural network for handwritten digit recognition on three quantum platforms: an IBM superconducting computer and two ion traps controlled by microwaves or lasers. They intentionally added randomness through quantum measurements during network operation, harnessing the inherent unpredictability of quantum superposition. On all three platforms, a moderate dose of quantum noise improved accuracy, while too much degraded it. For specific images that the network consistently misidentified, adding the right amount of randomness led to correct recognition almost always on the ion trap and about 90% of the time on the IBM machine. The researchers found that real quantum hardware outperformed simulations, attributing this to additional beneficial hardware noise. They plan to introduce quantum entanglement into the network next.