Neural-network Generated Quantum State Can Mitigate the Barren Plateau in Variational Quantum Circuits

TL;DR

The paper addresses barren plateaus in variational quantum circuits by proposing neural-network-generated quantum states (NGQS) to facilitate state preparation and alleviate optimization landscape flatness. It extends variational circuits to random configurations and jointly trains neural networks to emit complex quantum amplitudes, evaluating a cost that reflects the final state's nonzero probability. Empirical results show NGQS dramatically mitigates barren plateaus compared to random quantum states (RQS), with CNN- and larger-network architectures offering the strongest gains. The approach demonstrates a practical route to enhance variational quantum algorithms by integrating neural-state generation, especially for deeper circuits.

Abstract

We find that using neural networks to generate quantum states can effectively alleviate the barren plateau phenomenon present in random variational quantum circuits.

Figures (2)

• Figure 1: Illustration of NGQS models. In the NGQS model, the optimization step adjusts both $\alpha$ and the neural network, while the RQS model optimizes only $\alpha$. In the random quantum circuit, the box component corresponds to the defined block.
• Figure 2: Subfigure (a) depicting the average loss during training for ten models with different numbers of qubits; Subfigure (b) and Subfigure (c) depicting the average loss versus iterations during training at different depths.