Single-channel blind separation using L1-sparse complex non-negative matrix factorization for acoustic signals.

An innovative method of single-channel blind source separation is proposed. The proposed method is a complex-valued non-negative matrix factorization with probabilistically optimal L1-norm sparsity. This preserves the phase information of the source signals and enforces the inherent structures of the temporal codes to be optimally sparse, thus resulting in more meaningful parts factorization. An efficient algorithm with closed-form expression to compute the parameters of the model including the sparsity has been developed. Real-time acoustic mixtures recorded from a single-channel are used to verify the effectiveness of the proposed method.

Underdetermined reverberant acoustic source separation using weighted full-rank nonnegative tensor models.

In this paper, a fusion of K models of full-rank weighted nonnegative tensor factor two-dimensional deconvolution (K-wNTF2D) is proposed to separate the acoustic sources that have been mixed in an underdetermined reverberant environment. The model is adapted in an unsupervised manner under the hybrid framework of the generalized expectation maximization and multiplicative update algorithms. The derivation of the algorithm and the development of proposed full-rank K-wNTF2D will be shown. The algorithm also encodes a set of variable sparsity parameters derived from Gibbs distribution into the K-wNTF2D model. This optimizes each sub-model in K-wNTF2D with the required sparsity to model the time-varying variances of the sources in the spectrogram. In addition, an initialization method is proposed to initialize the parameters in the K-wNTF2D. Experimental results on the underdetermined reverberant mixing environment have shown that the proposed algorithm is effective at separating the mixture with an average signal-to-distortion ratio of 3 dB.

Single-channel blind separation using pseudo-stereo mixture and complex 2-D histogram.

A novel single-channel blind source separation (SCBSS) algorithm is presented. The proposed algorithm yields at least three benefits of the SCBSS solution: 1) resemblance of a stereo signal concept given by one microphone; 2) independent of initialization and a priori knowledge of the sources; and 3) it does not require iterative optimization. The separation process consists of two steps: 1) estimation of source characteristics, where the source signals are modeled by the autoregressive process and 2) construction of masks using only the single-channel mixture. A new pseudo-stereo mixture is formulated by weighting and time-shifting the original single-channel mixture. This creates an artificial mixing system whose parameters will be estimated through our proposed weighted complex 2-D histogram. In this paper, we derive the separability of the proposed mixture model. Conditions required for unique mask construction based on maximum likelihood are also identified. Finally, experimental testing on both synthetic and real-audio sources is conducted to verify that the proposed algorithm yields superior performance and is computationally very fast compared with existing methods.

Variational regularized 2-D nonnegative matrix factorization.

A novel approach for adaptive regularization of 2-D nonnegative matrix factorization is presented. The proposed matrix factorization is developed under the framework of maximum a posteriori probability and is adaptively fine-tuned using the variational approach. The method enables: (1) a generalized criterion for variable sparseness to be imposed onto the solution; and (2) prior information to be explicitly incorporated into the basis features. The method is computationally efficient and has been demonstrated on two applications, that is, extracting features from image and separating single channel source mixture. In addition, it is shown that the basis features of an information-bearing matrix can be extracted more efficiently using the proposed regularized priors. Experimental tests have been rigorously conducted to verify the efficacy of the proposed method.

Nonlinear signal separation for multinonlinearity constrained mixing model.

In this letter, a new type of nonlinear mixture is derived and developed into a multinonlinearity constrained mixing model. The proposed signal separation solution integrates the Theory of Series Reversion with a polynomial neural network whereby the hidden neurons are spanned by a set of mutually reversed activation functions. Simulations have been undertaken to support the theory of the proposed scheme and the results indicate promising performance.

Prediction of nodal spread of breast cancer by using artificial neural network-based analyses of S100A4, nm23 and steroid receptor expression.

The expression of tumour promoter gene S100A4, metastasis suppressor gene nm23, oestrogen and progesterone receptors, and tumour grade and size have been investigated for their potential to predict breast cancer progression. The molecular and cellular data have been analysed using artificial neural networks to determine the potential of these markers to predict the presence of metastatic tumour in the regional lymph nodes. This study shows that tumour grade and size are poor predictors. The relative expression of S100A4 and nm23 genes is the single most effective predictor of nodal status. Inclusion of oestrogen- and progesterone-receptor status with tumour grade and size markers improves prediction; however, there may be some overlap between steroid receptors and molecular markers. This study also underscores the power of artificial neural network techniques to predict the potential of primary breast cancers to spread to axillary lymph nodes. This could aid the clinician in determining whether invasive procedures of axially node dissection can be obviated and whether conservative forms of treatment might be appropriate in the management of the patient.

Oestrogen and progesterone receptor expression influences DNA ploidy and the proliferation potential of breast cancer cells.

The influence of oestrogen (ER) and progesterone (PgR) receptor has not been investigated in relation to DNA ploidy and cell proliferation. Here we have investigated a series of 46 breast cancer fine-needle aspirates in order to define the prognostic value of ER/PgR and possible correlations between DNA ploidy, size of the S-phase fraction (SPF) and cell cycle distribution features measured by image cytometry (ICM) and ER/PgR status of the primary tumours. The breast cancers were grouped into ER+/PgR+, ER+/PgR- and ER-/PgR-. Inter-group comparisons were made of DNA ploidy, SPF and the pattern of cell cycle distribution defined by the G0G1/G2M ratio in order to determine the influence of ER and PgR expression on the respective cell features. Our studies suggest that ER and PgR exert differential effects on the cell features examined. This study also analysed the possibility of predicting nodal involvement and 5-year disease-free survival by combining ER/PgR status with DNA ploidy, SPF and G0G1/G2M ratios. A high degree of accuracy was achieved for predicting both nodal involvement and 5-year disease-free survival. These findings suggest that a combination of ER/PgR status with DNA ploidy, SPF and cell cycle distribution provide a powerful marker for disease prognosis.

Analog design of a new neural network for optical character recognition.

An electronic circuit is presented for a new type of neural network, which gives a recognition rate of over 100 kHz. The network is used to classify handwritten numerals, presented as Fourier and wavelet descriptors, and has been shown to train far quicker than the popular backpropagation network while maintaining classification accuracy.

The DSFPN, a new neural network for optical character recognition.

A new type of neural network for recognition tasks is presented in this paper. The network, called the dynamic supervised forward-propagation network (DSFPN), is based on the forward only version of the counterpropagation network (CPN). The DSFPN, trains using a supervised algorithm and can grow dynamically during training, allowing subclasses in the training data to be learnt in an unsupervised manner. It is shown to train in times comparable to the CPN while giving better classification accuracies than the popular backpropagation network. Both Fourier descriptors and wavelet descriptors are used for image preprocessing and the wavelets are proven to give a far better performance.