Optimization, trainability & noise
Four detailed lessons, mathematical definitions, worked examples and limitations.
Explore the lessons
Concept 21Barren plateausGradients of certain deep or highly expressive random circuits may vanish exponentially with qubit number.Concept 22Local costs and layerwise trainingLocal observables, problem-inspired ansätze and incremental depth may improve trainability.Concept 23Shot noise and measurement budgetsEstimating observables requires repeated measurements; uncertainty shrinks as the inverse square root of shots.Concept 24Hardware noise and error mitigationReadout errors, decoherence and gate infidelity perturb the learned model and its gradients.