Majoranas Inside: Hybrid quantum-classical neural network with a fermionic layer

In this work, we introduce a hybrid quantum-classical neural network whose quantum layer is based on the paradigm of fermionic quantum computing. This new model is designed to enhance the capabilities of fermionic machine learning (FermiML) introduced in [arXiv:2404.19032]. Since fermionic quantum circuits are efficiently simulable classically, our work introduces a scalable benchmark for hybrid quantum-classical machine learning models. We conduct a systematic evaluation of this model on supervised tasks such as classification and regression, evaluating its performance against state-of-the-art hybrid quantum-classical neural networks in the literature. Our findings indicate that the proposed model performs competitively across multiple metrics, often outperforming existing architectures. Additionally, in the context of previously proposed hybrid neural networks, we observe that the inclusion of a quantum layer offers no significant benefit to learning on most datasets. In contrast, incorporating fermionic layers tends to enhance accuracy and scalability, suggesting their potential value in hybrid architectures.