Tensorized Neural Layer Decomposition for 2-D DOA Estimation

Existing matrix-based neural network for direction-of-arrival (DOA) estimation has to train a large amount of parameters proportional to the length of vectorized signal statistics, resulting in a heavy system overload. To address the problem, a tensorized neural layer decomposition-based neural network is proposed for 2-D DOA estimation. In particular, the covariance tensor of tensor signals is propagated to hidden state tensors. The feedforward propagation is formulated as an inverse Tucker decomposition, such that parameters in the tensorized neural layers are compressed into inverse Tucker factors. Accordingly, the tensorized backpropagation procedure is designed for network training. It is proved that the number of parameters is significantly reduced, which leads to a faster training process. Simulation results demonstrate that the proposed method reduces the number of trained parameters by more than 122,000 times compared to the matrix-based neural network while maintaining a moderate accuracy.