Layerwise learning for quantum neural networks

Abstract With the increased focus on quantum circuit learning for near-term applications devices, in conjunction with unique challenges presented by cost function landscapes of parametrized circuits, strategies effective training are becoming increasingly important. In order to ameliorate some these challenges, we investigate a layerwise strategy circuits. The depth is incrementally grown during optimization, and only subsets parameters updated each step. We show that when considering sampling noise, this can help avoid problem barren plateaus error surface due low number trained one step, larger magnitude gradients compared full circuit. These properties make our algorithm preferable execution noisy intermediate-scale devices. demonstrate approach an image-classification task handwritten digits, attains 8% lower generalization average comparison standard schemes circuits same size. Additionally, percentage runs reach test errors up 40% circuit, which susceptible creeping onto plateau training.