Remote DoA Estimation via Subspace-Oriented Deep-Learning-Aided Vector Quantization

Remote inference, where sensors acquire data and transmit compressed features to a remote server for inference, plays a key role in numerous applications. A common remote inference task is direction of arrival (DoA) estimation, where the sensed data is used for localizing multiple sources. Traditional methods require the sensor to downstream raw wideband data, leading to increased latency and spectral inefficiency. In this work, we propose Remote SubspaceNet, a deep neural network (DNN)-aided remote inference framework that enables interpretable, low-latency, and progressively refined DoA estimation at the server. Remote SubspaceNet integrates DNN-based feature extraction with learned vector quantization and subspace-based inference techniques, learning a single quantization codebook that supports successively refined DoA recovery. We demonstrate that Remote SubspaceNet accurately estimates DoAs across varying bit budgets, significantly reducing communication latency while maintaining interpretability and robustness.