Hidden Markov models as priors for regularized nonnegative matrix factorization in single-channel source separation

In this paper we explore avenues for improving the reliability of dimensionality reduction methods such as Non-Negative Matrix Factorization (NMF) as interpretive exploratory data analysis tools. We first explore the difficulties of the optimization problem underlying NMF, showing for the first time that non-trivial NMF solutions always exist and that the optimization problem is actually convex, by using the theory of Completely Positive Factorization. We subsequently explore four novel approaches to finding globally-optimal NMF solutions using various ideas from convex optimization. We then develop a new method, isometric NMF (isoNMF), which preserves non-negativity while also providing an isometric embedding, simultaneously achieving two properties which are helpful for interpretation. Though it results in a more difficult optimization problem, we show experimentally that the resulting method is scalable and even achieves more compact spectra than standard NMF.

The objective of single-channel source separation is to accurately recover source signals from mixtures. Non-negative matrix factorization (NMF) is a popular approach for this task, yet previous NMF approaches have not optimized directly this objective, despite some efforts in this direction. Our paper introduces discriminative training of the NMF basis functions such that, given the coefficients obtained on a mixture, a desired source is optimally recovered. We approach this optimization by generalizing the model to have separate analysis and reconstruction basis functions. This generalization frees us to optimize reconstruction objectives that incorporate the filtering step and SNR performance criteria. A novel multiplicative update algorithm is presented for the optimization of the reconstruction basis functions according to the proposed discriminative objective functions. Results on the 2nd CHiME Speech Separation and Recognition Challenge task indicate significant gains in source-to-distortion ratio with respect to sparse NMF, exemplar-based NMF, as well as a previously proposed discriminative NMF criterion.

A solution for separating speech from music signal as a single channel source separation is Non-negative Matrix Factorization (NMF). In this approach spectrogram of each source signal is factorized as multiplication of two matrices which are known as basis and weight matrices. To achieve proper estimation of signal spectrogram, weight and basis matrices are updated iteratively. To estimate distance between signal and its estimation a cost function is used usually. Different cost functions have been introduced based on Kullback-Leibler (KL) and Itakura-Saito (IS) divergences. IS divergence is scale-invariant and so it is suitable for the conditions in which the coefficients of signal have a large dynamic range, for example in music short-term spectra. Based on this IS property, in this paper, we propose to use IS divergence as cost function of NMF in the training stage for music and on the other hand we suggest to use KL divergence as NMF cost function in the training stage for speech. Moreover, in the decomposition stage, we propose to use a linear combination of these two divergences in addition to a regularization term which considers temporal continuity information as a prior knowledge. Experimental results on one hour of speech and music, shows a good trade-off between signal to inference ratio (SIR) of speech and music in comparison to conventional NMF methods.

Hidden Markov Models (HMMs) and associated Markov modulated time series are widely used for estimation and control in e.g. robotics, econometrics and bioinformatics. In this paper, we modify and extend a recently proposed approach in the machine learning literature that uses the method of moments and a Non-Negative Matrix Factorization (NNMF) to estimate the parameters of an HMM. In general, the method aims to solve a constrained non-convex optimization problem. In this paper, it is shown that if the observation probabilities of the HMM are known, then estimating the transition probabilities reduces to a convex optimization problem. Three recursive algorithms are proposed for estimating the transition probabilities of the underlying Markov chain, one of which employs a generalization of the Pythagorean trigonometric identity to recast the problem into a non-constrained optimization problem. Numerical examples are presented to illustrate how these algorithms can track slowly time-varying transition probabilities.

In recent years, nonnegative matrix factorization (NMF) with volume regularization has been shown to be a powerful identifiable model; for example for hyperspectral unmixing, document classification, community detection and hidden Markov models. In this paper, we show that minimum-volume NMF (min-vol NMF) can also be used when the basis matrix is rank deficient, which is a reasonable scenario for some real-world NMF problems (e.g., for unmixing multispectral images). We propose an alternating fast projected gradient method for min-vol NMF and illustrate its use on rank-deficient NMF problems; namely a synthetic data set and a multispectral image.

In this work, we will consider the dimension reduction of the set of time series, such as economic data, to find the meaningful basis vector for the set of data, and indicate which data use which basis vector. Usually each of the time series is analyzed independently in economics but here we will analyze the set of time series simultaneously. Since some of the economic data are measured as positive values and we want to decompose them as a mixture of the parts, we will apply non-negative matrix factorization to the economic data. Non-negative matrix factorization can compress dimensions by approximating a non-negative matrix with the product of two non-negative matrices. The two non-negative matrices are called the coefficient matrix and the basis matrix, and the basis matrix can be considered as a dimensionally compressed matrix. If the standard non-negative matrix factorization is used for economic data, the basis matrix may not be smooth. We think that the basis vectors should be smooth except a few special economical incidents. In the proposed method, a Graph-based non-negative matrix factorization is introduced to regularize the basis matrix of the time series. A path graph for representing the time series of economic data is incorporated into the non-negative matrix factorization as regularization. As a result, basis vectors that maintains the time series of economic data are decomposed. Furthermore, we propose to introduce a sparsity in the non-negative matrix factorization. Traditionally, the sparsity incorporated into non-negative matrix factorization has been used for basis vectors. However, the proposed method introduces the sparsity for coefficient vectors. Thus the proposed method, which simultaneously incorporates the sparsity for the coefficient vectors and the smoothness for the basis vectors, can extract the smooth basis vectors and the original economical data are approximated as the weighted sum of the few bases vectors. This allows us to discover economic trends and the best-fit trends for each data at the same time.

Guitar string separation is a novel and complicated task. Guitar notes are not pure steady-state signals, hence, we hypothesize that neither Non-Negative Matrix Factorization (NMF) nor Non-Negative Matrix Factor Deconvolution (NMFD) are optimal for separating them. Therefore, we separate steady-state and transient parts using Harmonic-Percussive Separation (HPS) as a preprocessing step. Then, we use NMF for factorizing the harmonic part and NMFD for deconvolving the percussive part. We make use of a hexaphonic guitar dataset which allows for objective evaluation. In addition, we compare several types of time-frequency mask and introduce an intuitive way to combine a binary mask with a ratio mask. We show that the HPS mask type has an effect on source estimation. Our proposed method achieved results comparable to NMF without HPS. Finally, we show that the optimal mask at the final separation stage depends on the estimation algorithm.

Nonnegative Matrix Factorization (NMF) is valuable in many applications of blind source separation, signal processing and machine learning. A number of algorithms that can infer nonnegative latent factors have been developed, but most of these assume a specific noise kernel. This is insufficient to deal with complex noise in real scenarios. In this paper, we present a hierarchical Dirichlet process nonnegative matrix factorization (DPNMF) model in which the Gaussian mixture model is used to approximate the complex noise distribution. Moreover, the model is cast in the nonparametric Bayesian framework by using Dirichlet process mixture to infer the necessary number of Gaussian components. We derive a mean-field variational inference algorithm for the proposed nonparametric Bayesian model. We first test the model on synthetic data sets contaminated by Gaussian, sparse and mixed noise. We then apply it to extract muscle synergies from the electromyographic (EMG) signal and to select discriminative features for motor imagery single-trial electroencephalogram (EEG) classification. Experimental results demonstrate that DPNMF performs better in extracting the latent nonnegative factors in comparison with state-of-the-art methods.

Traditional Nonnegative Matrix Factorization (NMF) is a linear and unsupervised algorithm. This would limit the classification power of NMF for the complicated data. To overcome the above limitations of NMF, this paper proposes a novel supervised and nonlinear NMF algorithm based on kernel theory and discriminant analysis. We incorporate the class label information into the decomposition of NMF in the Reproducing Kernel Hilbert Space (RKHS). A new iterative algorithm for NMF is derived and the objective function is non-increasing under the update rules. The proposed method is evaluated on the ORL and Yale face databases. The experimental results demonstrate the the proposed method is superior to the state of-the-art algorithms.

Non-negative matrix factorization for speech/music separation using source dependent decomposition rank, temporal continuity term and filtering

The Baum-Welch algorithm together with its derivatives and variations has been the main technique for learning hidden Markov models (HMMs) from observational data. We present an HMM learning algorithm based on the nonnegative matrix factorization (NMF) of higher order Markovian statistics that is structurally different from the Baum-Welch and its associated approaches. The described algorithm supports estimation of the number of recurrent states of an HMM and iterates the NMF algorithm to improve the learned HMM parameters. Numerical examples are provided as well.

Nonnegative matrix factorization (NMF) has been increasingly investigated for data analysis and dimension-reduction. To tackle large-scale data, several online techniques for NMF have been introduced recently. So far, the online NMF has been limited to the linear model. This paper develops an online version of the nonlinear kernel-based NMF, where the decomposition is performed in the feature space. Taking the advantage of the stochastic gradient descent and the mini-batch scheme, the proposed method has a fixed, tractable complexity independent of the increasing samples number. We derive the multiplicative update rules of the general form, and describe in detail the case of the Gaussian kernel. The effectiveness of the proposed method is validated on unmixing hyperspectral images, compared with the state-of-the-art online NMF methods.

In this work we consider the problem of Hidden Markov Models (HMM) training. This problem can be considered as a global optimization problem and we focus our study on the Particle Swarm Optimization (PSO) algorithm. To take advantage of the search strategy adopted by PSO, we need to modify the HMM's search space. Moreover, we introduce a local search technique from the field of HMMs and that is known as the Baum–Welch algorithm. A parameter study is then presented to evaluate the importance of several parameters of PSO on artificial data and natural data extracted from images.

In this paper, we propose a novel scheme to supervised nonnegative matrix factorization (NMF). We formulate the supervised NMF as a sparse optimization problem assuming the availability of a set of basis vectors, some of which are irrelevant to a given matrix to be decomposed. The proposed scheme is presented in the context of music transcription and musical instrument recognition. In addition to the nonnegativity constraint, we introduce three regularization terms: (i) a block ℓ <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sub> norm to select relevant basis vectors, and (ii) a temporal-continuity term plus the popular ℓ <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sub> norm to estimate correct activation vectors. We present a state-of-the-art convex-analytic iterative solver which ensures global convergence. The number of basis vectors to be actively used is obtained as a consequence of optimization. Simulation results show the efficacy of the proposed scheme both in the case of perfect/imperfect basis matrices.

In this study, single-channel speech source separation is carried out to separate the speech from the background music, which degrades the speech recognition performance especially in broadcast news transcription systems. Since the separation is done using single observation of the source signals, the sources have to be previously modeled using training data. Non-negative Matrix Factorization (NMF) methods are used to model the sources. In order to model the source signals, different training data sets, which contain different music and speech data, are created and the effect of the training data sets are analyzed in this study. The performances of the methods are measured not only using separation performance measure but also with speech recognition performance measures.