A novel pre-training method is proposed to improve deep-neural-networks (DNN) and long-short-term-memory (LSTM) performance, and reduce the local minimum problem for speech enhancement. We propose initializing the last layer weights of DNN and LSTM by Non-Negative-Matrix-Factorization (NMF) basis transposed values instead of random weights. Due to its ability to extract speech features even in presence of non-stationary noises, NMF is faster and more successful than previous pre-training methods for network convergence. Using NMF basis matrix in the first layer along with another pre-training method is also proposed. To achieve better results, we further propose training individual models for each noise type based on a noise classification strategy. The evaluation of the proposed method on TIMIT data shows that it outperforms the baselines significantly in terms of perceptual-evaluation-of-speech-quality (PESQ) and other objective measures. Our method outperforms the baselines in terms of PESQ up to 0.17, with an improvement percentage of 3.4%.

Background and Objectives: Community detection is a critical problem in ‎investigating complex networks. Community detection based on ‎modularity/general modularity density are the popular methods with the ‎advantage of using complex network features and the disadvantage of ‎being NP-hard problem for clustering. Moreover, Non-negative matrix ‎factorization (NMF)-based community detection methods are a family of ‎community detection tools that utilize network topology; but most of ‎them cannot thoroughly exploit network features. In this paper, a hybrid ‎NMF-based community detection infrastructure is developed, including ‎modularity/ general modularity density as more comprehensive indices of ‎networks. The proposed infrastructure enables to solve the challenges of ‎combining the NMF method with modularity/general modularity density ‎criteria and improves the community detection methods for complex ‎networks.‎Methods: First, new representations, similar to the model of symmetric ‎NMF, are derived for the model of community detection based on ‎modularity/general modularity density. Next, these indices are ‎innovatively augmented to the proposed hybrid NMF-based model as two ‎novel models called ‘general modularity density NMF (GMDNMF) and ‎mixed modularity NMF (MMNMF)’. In order to solve these two NP-hard ‎problems, two iterative optimization algorithms are developed.‎Results: it is proved that the modularity/general modularity density-‎based community detection can be consistently represented in the form ‎of SNMF-based community detection. The performances of the proposed ‎models are verified on various artificial and real-world networks of ‎different sizes. It is shown that MMNMF and GMDNMF models ‎outperform other community detection methods. Moreover, the ‎GMDNMF model has better performance with higher computational ‎complexity compared to the MMNMF model.‎Conclusion: The results show that the proposed MMNMF model improves ‎the performance of community detection based on NMF by employing ‎the modularity index as a network feature for the NMF model, and the ‎proposed GMDNMF model enhances NMF-based community detection by ‎using the general modularity density index.‎

Non-negative Matrix Factorization (NMF) is a part-based image representation method. It comes from the intuitive idea that entire face image can be constructed by combining several parts. In this paper, we propose a framework for face recognition by finding localized, part-based representations, denoted “ Iterative weighted non-smooth non-negative matrix factorization” (IWNS-NMF). A new cost function is proposed in order to incorporate sparsity which is controlled by a specific parameter and weights of feature coefficients. This method extracts highly localized patterns, which generally improves the capability of face recognition. After extracting patterns by IWNS-NMF, we use principle component analysis to reduce dimension for classification by linear SVM. The Recognition rates on ORL, YALE and JAFFE datasets were 97. 5, 93. 33 and 87. 8%, respectively. Comparisons to the related methods in the literature indicate that the proposed IWNS-NMF method achieves higher face recognition performance than NMF, NS-NMF, Local NMF and SNMF.

Nonnegative matrix factorization (NMF) is a common method in data mining that have been used in different applications as a dimension reduction, classification or clustering method. Methods in alternating least square (ALS) approach usually used to solve this non-convex minimization problem. At each step of ALS algorithms two convex least square problems should be solved, which causes high computational cost. In this paper, based on the properties of norms and orthogonal transformations we propose a framework to project NMF’ s convex sub-problems to smaller problems. This projection reduces the time of finding NMF factors. Also every method on ALS class can be used with our proposed framework.

Introduction: Non-invasive Fluorescent Reflectance Imaging (FRI) is used for accessing physiological and molecular processes in biological media. The aim of this article is to separate the overlapping emission spectra of quantum dots within tissue-equivalent phantom using SVD, Jacobi SVD, and NMF methods in the FRI mode.Materials and Methods: In this article, a tissue-like phantom and an optical setup in reflectance mode were developed. The algorithm of multispectral imaging method was then written in Matlab environment. The setup included the diode-pumped solid-state lasers at 479 nm, 533 nm, and 798 nm, achromatic telescopic, mirror, high pass and low pass filters, and EMCCD camera. The FRI images were acquired by a CCD camera using band pass filter centered at 600 nm and high pass max at 615 nm for the first region and high pass filter max at 810 nm for the second region. The SVD and Jacobi SVD algorithms were written in Matlab environment and compared with a Non-negative Matrix Factorization (NMF) and applied to the obtained images.Results: PSNR, SNR, CNR of SVD, and NMF methods were obtained as 39 dB, 30.1 dB, and 0.7 dB, respectively. The results showed that the difference of Jacobi SVD PSNR with PSNR of NMF and modified NMF algorithm was significant (p<0.0001). The statistical results showed that the Jacobi SVD was more accurate than modified NMF.Conclusion: In this study, the Jacobi SVD was introduced as a powerful method for obtaining the unmixed FRI images. An experimental evaluation of the algorithm will be done in the near future.

Community detection is a significant issue in extracting valuable information and ‎understanding complex network structures. Non-negative matrix factorization (NMF) methods ‎are the most remarkable topics in community detection. The modularized tri-factor NMF ‎‎(MtriNMF) method was proposed as a new class of NMF methods that combines the modularized ‎information with tri-factor NMF. It has high computational complexity due to its dependence on ‎the choice of the initial value of the parameter and the number of communities (c). In other ‎words, the MtriNMF method should search among different c candidates to find correct c. In this ‎paper, a novel hybrid adaptive MtriNMF (HamtriNMF) method is proposed to improve the ‎performance of MtriNMF and reduce the computational complexity efficiently. In the proposed ‎method, computational complexity reduction is made by selecting the right c candidates and ‎tuning parameter. For this purpose, a hybrid algorithm including singular value decomposition ‎‎(SVD) and relative eigenvalue gap (REG) algorithms is suggested to estimate the set of c ‎candidates. Next, the TpmtriNMF model is proposed to improve the performance of community ‎detection via employing a self-tuning β parameter. Moreover, experimental results confirm the ‎efficiency of the HamtriNMF method with respect to other reference methods on artificial and ‎real-world networks.‎