Enhanced Quantum Neural Network Design for Nonlinear Training

Quantum neural networks (QNNs) offer higher-dimensional spaces than classical neural networks, giving them potential to capture deeper relationships in data. In practice, however, a QNN’s expressivity is often bounded by the nonlinearity of its feature map (the data-embedding circuit). Superposed Parameterized Quantum Circuits (SPQCs) address this weakness by introducing a single nonlinear term after the feature map, enabling nonlinear transformations during training. We propose an enhanced SPQC with an added qubit register to introduce multiple nonlinear terms. The number of added terms scales linearly with the size of this register, yielding a controllable increase in model expressivity. We present our circuit design and report empirical results on classification tasks, including learning curves and decision-boundary visualizations.