Quantum computers may one day enable us to solve many classically intractable problems. However, current devices are too noisy to be used for full fault-tolerant quantum computation. Near-term quantum algorithms attempt to extract useful performance from these devices and are a crucial bench-marking tool for device comparison and development.
In this thesis we introduce several, primarily machine learning-based, methods for reducing the quantum resources required in various stages of near-term quantum algorithms. First, we introduce a tensor network approach for designing low-depth and parameter-efficient ansatz circuits for modest sized variational algorithms. Next, we turn to the implementation of these variational algorithms and show that a Bayesian optimizer using a quantum kernel-based surrogate model greatly reduces the number of circuits submitted to the device.
We then consider the problem of efficiently characterizing an algorithm's output quantum state, introducing a neural network architecture which can be used to learn and sample from the outcome distributions of a state in arbitrary local measurement bases.
Finally, we demonstrate how a simple bit-flipping strategy can be used to simplify the effective readout errors experienced on a noisy device. As a result, far fewer calibration measurements are needed to mitigate the effective readout errors in post-processing, enabling more accurate quantitative results to be obtained from near-term algorithms.