Large neural networks have aroused impressive progress in various real world applications. However, the expensive storage and computational resources requirement for running deep networks make them problematic to be deployed on mobile devices. Recently, matrix and tensor decompositions have been employed for compressing neural networks. In this paper, we develop a simultaneous tensor decomposition technique for network optimization. The shared network structure is first discussed. Sometimes, not only the structure but also the parameters are shared to form a compressed model at the expense of degraded performance. This indicates that the weight tensors between layers within one network contain both identical components and independent components. To utilize this characteristic, two new coupled tensor train decompositions are developed for fully and partly structure sharing cases, and an alternating optimization approach is proposed for low rank tensor computation. Finally, we restore the performance of the neural network model by fine-tuning. The compression ratio of the devised approach can then be calculated. Experimental results are also included to demonstrate the benefits of our algorithm for both applications of image reconstruction and classification, using the well known datasets such as Cifar-10/Cifar-100 and ImageNet and widely used networks such as ResNet. Comparing to the state-of-the-art independent matrix and tensor decomposition based methods, our model can obtain a better network performance under the same compression ratio.

Synchronistical localization, separation, and reconstruction for multiple sound sources are usually necessary in various situations, such as in conference rooms, living rooms, and supermarkets. To improve the intelligibility of speech signals, the application of deep neural networks (DNNs) has achieved considerable success in the area of time-domain signal separation and reconstruction. In this paper, we propose a hybrid microphone array signal processing approach for the nearfield scenario that combines the beamforming technique and DNN. Using this method, the challenge of identifying both the sound source location and content can be overcome. Moreover, the use of a sequenced virtual sound field reconstruction process enables the proposed approach to be quite suitable for a sound field which contains a dominant, stronger sound source and masked, weaker sound sources. Using this strategy, all traceable, mainly sound, sources can be discovered by loops in a given sound field. The operational duration and accuracy of localization are further improved by substituting the broadband weighted multiple signal classification (BW-MUSIC) method for the conventional delay-and-sum (DAS) beamforming algorithm. The effectiveness of the proposed method for localizing and reconstructing speech signals was validated by simulations and experiments with promising results. The localization results were accurate, while the similarity and correlation between the reconstructed and original signals was high.

Constraint by the physical geometry, the lower and upper frequency bound and the scale of the scanning area of a microphone array are limited. Owing to its movable feature, for the service robots, achieving a wider working frequency range with a global view requires a virtually larger and denser array, which can be realised using non-synchronous measurements beamforming with a movable microphone array prototype. However, even when using the state-of-the-art method, it is challenging to localise multiple broadband sources, owing to the difficulty in selecting an appropriate operating frequency without any prior information about the target signal. Therefore, this letter proposes a tensor-completion-based non-synchronous measurements method for broadband multiple-sound-source localisation. The tensor data structure of the broadband signal is analysed, and an alternating direction method based on multiplier optimisation with a tensor multi-norm constraint is proposed. This algorithm can provide a sound map with a distinct global view of three different speech signal sources with high accuracy. Compared with the matrix-based optimisation method, the proposed method can significantly reduce the mean square error of the estimated source location.

This letter addresses the problem of tensor completion. The properties of the tensor tubal rank (TTR) and tensor Kronecker rank are first discussed, and then a novel generalized tubal Kronecker decomposition together with a new tensor rank referred to as generalized tensor tubal rank (GTTR) are defined. It is shown that the GTTR is suitable for revealing both the Kronecker and tubal structures of a tensor. The general tensor completion idea is then presented following the procedure of alternate projection between tensor rank minimization and Frobenius-norm optimization. Furthermore, the GTTR minimization is relaxed to the problem of generalized tensor nuclear norm (TNN) minimization, and two solutions are derived. The first one is based on the idea of combining all generalized TNNs as a weighted sum, while the second one employs the alternate cancelation scheme. Experiments are also carried out using both simulated data and real datasets for comparison of the proposed and the state-of-the-art approaches.