Introduction
As the telecommunications industry pivots toward 6G, Integrated Sensing and Communications (ISAC) has emerged as a cornerstone technology. ISAC leverages existing Wi-Fi signals to perform environmental sensing—such as human pose recognition and gesture tracking—without the privacy risks associated with optical cameras. While Deep Neural Networks (DNNs) are the current standard for processing these signals, their high computational cost and large parameter counts pose significant challenges for deployment on edge devices. Quantum Machine Learning (QML) offers a promising alternative, potentially providing extreme efficiency through Quantum Neural Networks (QNNs).
The Challenge of Quantum Circuit Design
A primary hurdle in applying QML to real-world tasks is the "Ansatz" problem. An ansatz is the structural template of a quantum circuit, defining the arrangement of qubits, gates, and entanglement layers. Historically, designing an effective ansatz has been a manual, trial-and-error process. To bridge this gap, researchers from Mitsubishi Electric Research Laboratories (MERL) developed AutoAnsatz, an automated framework designed to optimize QNN architectures for Wi-Fi sensing.
The AutoAnsatz Framework
AutoAnsatz utilizes Bayesian Optimization, specifically the Tree-Parzen Estimator, to navigate a vast search space of quantum circuit configurations. The framework evaluates various ansatz types, including Matrix Product States (MPS) and Tree Tensor Networks (TTN), alongside different data embedding methods and hyperparameters like qubit count and layer depth.
To maintain efficiency during the search, the system employs Hyperband pruning. This technique identifies and terminates underperforming circuit designs early in the training phase, allowing the optimizer to concentrate resources on the most viable candidates.
Experimental Results and Efficiency
The framework was tested on a custom dataset of 60 GHz mmWave Wi-Fi signals. The task involved recognizing eight distinct human poses based on 36-dimensional beam Signal-to-Noise Ratio (SNR) measurements.
The results demonstrate a remarkable leap in efficiency. While a baseline DNN required approximately 35,000 parameters to achieve high accuracy, the AutoAnsatz-optimized QNNs reached comparable performance (>80% accuracy) with significantly fewer resources. Specifically, a 10-qubit Matrix Product State model achieved state-of-the-art results using only 54 parameters—a 650-fold reduction compared to the DNN.
Functional analysis of variance (fANOVA) revealed that the choice of ansatz is the most critical factor in model performance, outweighing other variables like the number of qubits. Furthermore, the study found that shallower circuits often performed better by avoiding "barren plateaus," a common issue in quantum optimization where gradients vanish.
Conclusion and Future Outlook
AutoQML demonstrates that automated design can unlock the potential of QML for practical sensing applications. By achieving high accuracy with thousands of times fewer parameters than classical models, QNNs present a viable path for low-power, high-performance sensing in future 6G networks. Future work will focus on validating these designs on physical Noisy Intermediate-Scale Quantum (NISQ) hardware and refining embedding techniques to better handle the complexities of real-world signal noise.
Summary updated