Parallel Data Processing in Quantum Machine Learning

Abstract

We propose a Quantum Machine Learning (QML) framework that applies the core design principle of quantum algorithms-superposition, oracle, and interference-to accelerate training. Building on the structural analogy between feature extraction in foundational quantum algorithms and parameter optimization in QML, we reformulate the training process to leverage quantum parallelism: all training samples are encoded into a quantum superposition, processed through a parameterized quantum circuit, and classified via an interferometer module that implements quantum interference across the dataset. This architectural reformulation reduces the theoretical complexity of loss function evaluation from $O(N^{2})$ in conventional QML training to $O(N)$, where $N$ is the dataset size. Numerical simulations on multiple binary and multi-class classification datasets (with up to $N=128$ samples) demonstrate that our method achieves classification accuracies comparable to conventional circuits while reducing the number of quantum circuit executions per cost function evaluation from $N$ to 1. This represents a near $N$-fold reduction in quantum overhead per training iteration, reducing the required circuit executions without loss of accuracy. These results highlight the potential of quantum algorithmic design principles as a scalable pathway to efficient QML implementations.