Entanglement-enabled advantage in distributed quantum machine learning

Large-scale quantum computing is limited by the qubit number and coherence time of devices in the NISQ era. The quantum internet provides a scalable framework by linking smaller processors, enabling distributed quantum computing across distance. However, communication latency poses a major obstacle, as the time required for information exchange can exceed the coherence time of qubits. In contrast, pre-shared entanglement can be established beforehand and used to reduce communication complexity between remote nodes. Here, we implement distributed quantum machine

learning (DQML) on two 4-qubit processors, varying the number of pre-shared Bell pairs to perform binary classification on 8-dimensional datasets. By analogy with the CHSH game, we show that entanglement enhances classification accuracy, with the gain depending on the data embedding and loss-function choice. On synthetic datasets, the accuracy improves up to 30%,

particularly for complex structures. Interestingly, we find that excessive entanglement could degrade accuracy, suggesting that an optimal level of entanglement is crucial to achieving the advantage. These results highlight the power of entanglement as a quantum channel for DQML, outlining a route toward practical distributed quantum computation beyond coherence-time limits.