Adding physical noise to quantum neural networks can improve their accuracy on certain datasets, as demonstrated on Quandela’s Simulator for hybrid quantum-classical neural networks

Researchers have discovered that intentionally adding physical noise to quantum neural networks can improve their accuracy on certain datasets. Using Quandela’s Perceval simulator and a complex genetic algorithm, the team injected a seven-parameter physical noise model into photonic hybrid quantum-classical neural networks, optimizing the noise itself for each dataset. This approach yielded accuracy gains of 0.82 percentage points on the Iris dataset and 1.45 percentage points on Digits, mirroring noise-injection techniques used in classical deep learning. However, the work also reveals a degradation of 1.21 percentage points on the MNIST dataset, demonstrating that the benefit of this noise-as-regularization technique is dataset-dependent and requires careful tuning. The ability to intentionally introduce imperfections into quantum systems to improve performance appears counterintuitive, yet recent work demonstrates this with photonic hybrid quantum-classical neural networks (PHQCNNs).…