QERGD+: Quantum-Enhanced Recurrent Gradient Descent with Shor-Inspired Periodicity Priors for Stabilizing Dynamic Neural Network Training
Year: 2025
Type: article
Abstract: This research introduces QERGD+, a hybrid quantum–classical algorithm for training recurrent neural networks (RNNs) and gated variants (GRUs, LSTMs) on long sequences without suffering from the vanishing/exploding gradient problem inherent to backpropagation through time (BPTT). The method replaces unstable recurrent-path gradients with exact, analytic parameter-shift derivatives computed on parameterized quantum circuits (PQCs) using Qiskit primitives. It augments training with a Shor-inspired, low-depth periodicity prior that biases learning toward stable cycles when periodic or quasi-periodic structure is present. The approach confines quantum resources to the recurrent bottleneck, remains NISQ-friendly, and maintains compatibility with large-scale classical architectures. Detailed pseudocode, Qiskit implementation patterns, complexity analysis, and empirical use-case guidance are provided.
Author: Lakshit Mathur
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W6949608740
FWCI:
Cites:
Cited by:
Related to: