Generalized quantum neural modules for high-dimensional local quantum feature learning

Abstract

Quantum convolution offers a natural inductive bias for extracting local structure from high-dimensional data, yet its practical reach is constrained by direct feature-wise encoding, which ties broader receptive fields to wider quantum registers. This raises a central question: can local quantum representations retain their predictive content when receptive-field growth is redirected from circuit width to sequential processing on a compact register? Generalized quantum neural modules address this question through incremental data upload, partitioning each receptive field into ordered batches, preserving a single evolving quantum state across successive encode–entangle stages, and reusing shared trainable transformations as the local window moves across the input. State-vector experiments on controlled three- and four-dimensional motif tensors, MNIST images, and selected UCF101 action videos show that moderate serialization substantially narrows the active register while preserving the predictive behaviour of direct encoding; more aggressive partitioning, by contrast, reveals a progressive trade-off between sequential depth and within-batch feature interaction, with mixed CRZ–CRX entanglement providing the most effective fixed-width configuration. Together, these results delineate a practical operating regime for high-dimensional local quantum learning and show that receptive-field scaling can be managed through a coordinated choice of upload granularity, entangling structure, information retention, and execution depth.