The Exploration on the Variety of Methods that Attempt to Reduce Barren Plateaus in Quantum Machine Learning, and the Proposal of a Noise-Reducing Ansatz as a Solution – A Literature Review

Abstract

Barren plateaus have notably decreased the ease of trainability of Quantum Machine Learning models. Barren plateaus are formed where the variability decreases exponentially in the graph of the cost function. The purpose of this literature review was to provide an up-to-date overview of solutions to reduce barren plateaus, by analyzing the effects of Quantum Convolutional Neural Networks (QCNN), novel perturbations in Variational Quantum Algorithms, Variational Quantum State Diagonalization, and the Expressiveness of Parametrization on reducing barren plateaus. The novel perturbations in step sizes in Quantum Natural Gradient proved to be more effective than the QCNN, which has been widely recognized as one of the most efficient ways to maintain variability in a cost function to avoid barren plateaus. Additionally, the use of an ansatz (a circuit with a predetermined geometry and parameterized gates) reduces the hardware noise of a quantum computer leading to reduce noise-induced barren plateaus. These results serve to encourage further research into novel perturbations in the steps of Quantum Natural Gradient and the Noise-Reducing Ansatz.