This work presents a study of a novel implementation of a Quantum Convolutional Neural Network (QCNN) aimed at binary classification of images from the Modified National Institute of Standards and Technology (MNIST) dataset. Inspired by previous QCNN and classical convolutional neural network (CNN) implementations, we employ a hierarchical partitioning approach to implement a QCNN circuit that can be approximated and simulated efficiently on a classical machine for large problems.
Initially, the original image is partitioned so that each process handles a smaller portion, encoded into independent states. These partitions then merge, creating states that contain information from both partitions while halving the number of processes.
This process is repeated until one process remains, at which point we reduce the dimensionality of the state until only a single qubit remains for measurement.
Using this approach, we can leverage multiple processes in parallel to simulate a large QCNN program without the exponential growth in hardware requirements as the number of qubits increases. In our work, we train a 128-qubit model, which cannot be executed on any classical supercomputer without this novel architecture.
We also explore the impact of this model architecture on prediction accuracy by training it to perform binary classification on the MNIST dataset and comparing it to a model without partitioning. Our initial findings indicate that partitioning images into smaller sub-images does not degrade model performance and may even improve it, likely due to alleviating the Barren plateaus issue during partitioning.
Blogger's Review: This research showcases an innovative hierarchical partitioning method, demonstrating the feasibility of quantum neural networks on classical computers while addressing hardware limitations in quantum computing. This approach not only enhances model scalability but may also lead to improved prediction accuracy, warranting further exploration and application.