A linear feedforward neural network with lateral feedback connections for blind source separation

Blind source separation (BSS) consists in processing a set of observed mixed signals to separate them into a set of original components. Most of the current blind separation methods assumes that the source signals are “as statistically independent as possible” given the observed data. In many real-world situations, however, this hypothesis does not hold. In order to cope with such signals, a first geometric method was proposed that separates statistically dependent signals, provided that they are nonnegative and locally orthogonal.This paper presents a new geometric method for the separation of nonnegative source signals which relies on a working assumption that is weaker than local orthogonality. The separation problem is expressed as the identification of relevant facets of the data cone. After a rigorous proof of the proposed method, the details of the separation algorithm are given. Experiments on signals from various origins clearly show the efficiency of the new procedure.

In order to overcome the problem of low accuracy and poor stability caused by the one-sided information blind separation analysis of the current communication channel interference suppression research methods, this paper proposes the optical communication network channel interference suppression method based on blind signal separation algorithm. The receiver unit, field programmable gate array (FPGA) unit, memory chip group and other modules are used to construct the channel signal acquisition system, and the collected signals are preprocessed by centralisation and whitening. Based on the signal preprocessing results, the blind separation of instantaneous linear mixed signals, blind separation of convolutional mixed signals and blind separation of sparse sources were comprehensively analysed to obtain the source signals and realise channel interference suppression of optical communication network. Experimental results show that this method has high interference precision, strong stability and good application prospect.

Human action recognition is a research hotspot in the field of ambient intelligence, serving as a foundation for the Internet of Healthcare Things (IoHT) and smart home wellness, with extensive application value. Currently, most research primarily focuses on the performance of single-modal approach for human action recognition. However, single-modal data cannot adequately capture the action characteristics of the human body. For instance, camera data lacks detailed micro-motion information, while radar data lacks visual appearance information. This limitation makes human action recognition systems susceptible to complex covariate influences. To address this issue, this paper proposes a Multi-Modal Fusion Model, M2FM, to accurately recognize various complex human action from millimeter-wave radar signals and video data. In the millimeter-wave radar branch, a LFNet network is constructed to capture richly hierarchical human action representation by extracting micro-Doppler feature and cadence velocity feature simultaneously. Specially, a Linear Feedforward Neural Network (LFNN) module is designed for modeling global-local feature of human action. In the video branch, a lightweight Transformer-based video analysis and action recognition network, STL-Former, is developed. Specially, an Agent downSampling (AS) attention module and a Linear Maxpooling (LM) attention module are designed to achieve context-aware downsampling with global receptive field and spatial vector sequence downsampling, respectively. Additionally, a multi-scale Linear attention (Litner) module is designed for modeling the spatio-temporal information of human action. Finally, the outputs from two branches are fused by a filter-based Dempster-Shafer theory. The experimental results on a self-collected Multi-modal Human Action Dataset, JH-MHAD, show that the M2FM model outperforms other advanced models, with an average recognition accuracy of 99.3% for eight different human actions, and it demonstrates strong adaptability under different conditions.

Independent component analysis (ICA) has shown much success in blind source separation and channel equalization. Its applications to remotely sensed images are investigated in recent years. Linear spectral mixture analysis (LSMA) has been widely used for subpixel detection and mixed pixel classification in remote sensing image processing and already studied in Chapters 3, 8–10, . It models the spectral signature of an image pixel vector as a linear mixture of spectral signatures of targets present in the image data and the target abundance fractions are assumed to be unknown, but nonrandom constants. This chapter combines these two approaches to one, called ICA-based linear spectral random mixture analysis (LSRMA), which describes an image pixel vector as a random process resulting from a random composition of multiple spectral signatures of distinct targets in the image data. It differs from LSMA in that the abundance fractions of the target signatures in LSRMA are considered to be unknown random independent signal sources. Two major advantages can be benefited from LSRMA. First, LSRMA does not require any prior knowledge of the targets used in the linear mixture model. Second and most importantly, LSRMA models each of target signatures as an independent random signal source so that the spectral variability of target signatures can be captured more effectively in a stochastic manner. The only required knowledge for LSRMA is the number of signal sources, p assumed to be present in the image data. This number is estimated by Neyman-Pearson detection theory-based eigen-thresholding methods to be investigated in Chapter 17.

This piece of research adopts an interdisciplinary conceptual approach that incorporates Artificial Neural Networks (ANN) with learning and cognitive sciences. Specifically, it considers modeling of associative memorization to introduce optimal analysis for development of reading brain performance. Herein, this brain performance simulated realistically using ANNs self-organized modeling paradigm. That namely: the Generalized Hebbian Algorithm (GHA), also known in the literature as Sanger's rule, is a linear feed forward neural network model for unsupervised learning with applications primarily in principal components analysis. Furthermore, it inspired by functioning of highly specialized biological neurons in reading brain based on the organization the brain's structures/substructures. In accordance with the prevailing concept of individual intrinsic characterized properties of highly specialized neurons. Presented models have been in close correspondence with set of neurons' performance for developing reading brain in a significant way. More specifically, herein, introduced model concerned with their important role played in carrying out cognitive reading brain function's outcomes. Accordingly, the cognitive goal for reading brain is to translate that seen word (orthographic word-from) into a spoken word (phonological word-form). In this context herein, the presented work illustrates via ANN simulation results: How ensembles of highly specialized neurons could be dynamically involved in performing associative memorization cognitive function for developing reading brain.

Blind Signal Separation II. Linear, Convolutive Combinations: II. Linear, Convolutive Combinations

Combining the advantages of local mean decomposition and variational Bayesian hidden Markov model, a blind source separation method for mechanical faults based on local mean decomposition and variational Bayesian hidden Markov model is proposed. In the proposed method, the non-stationary signals are decomposed into a series of production functions by the local mean decomposition, then the obtained production functions and the original observed signals are used to construct new observation signals, therefore the underdetermined blind source separation problem is transformed into the overdetermined blind source separation problem. Finally, the source signals are estimated by the variational Bayesian hidden Markov model method. The proposed method has been successfully applied to the blind separation of bearing faults, and the experiment results verify the effectiveness of the proposed method. The obtained outcome in this paper provides an effective separation method for mechanical fault diagnosis.

Most of the existing algorithms for blind sources separation have a limitation that sources are statistically independent. However, in many practical applications, the source signals are nonnegative and mutual statistically dependent signals. When the observations are nonnegative linear combinations of nonnegative sources, the correlation coefficients of the observations are larger than these of source signals. In this letter, a novel Nonnegative Matrix Factorization (NMF) algorithm with least correlated component constraints to blind separation of convolutive mixed sources is proposed. The algorithm relaxes the source independence assumption and has low-complexity algebraic computations. Simulation results on blind source separation including real face image data indicate that the sources can be successfully recovered with the algorithm.

Discusses the relationship between the mean square classifier (MSC) and the linear feedforward neural network classifier (LFNNC) and further studies the transformation properties of LFNNs under the supervision of outer-supervised signals. The obtained conclusions show that an LFNNC is equivalent to the cascade of many MSCs, and vice versa (only if the hidden units number is greater than the "bottleneck" limit). Therefore, we can cascade several MSCs to form a modular LFNNC which is completely equivalent to doing information fusion from different MSCs.

The paper presents a linear feedback neural network and an adaptive algorithm to achieve the blind signal separation under near-field situation. The convergence property of algorithm and stability of equilibrium state are discussed. Some simulations are provided.

The blind source separation problem is considered through the approach based on non-stationarity and coloration. In both cases, the sources are usually assumed to be Gaussian. In this paper, we extend previous works in order to handle sources drawn from the multivariate Student t-distribution. After studying the structure of the parameter manifold in this case, a new blind source separation criterion based on the log-likelihood of the considered distribution is proposed. To solve the resulting optimization problem, Riemannian optimization on the parameter manifold is leveraged. Practical expressions of the mathematical tools required by first order based Riemmanian optimization methods for this parameter manifold are derived to this end. The performance of the proposed method is illustrated on simulated data.

Blind extraction or separation statistically independent source signals from their linear mixtures have been well studied in the last two decades, which is realised by searching for local extrema of certain objective functions. In this paper, by employing nonparametric non-Gaussianity (NG) measure, a blind source separation/extraction (BSS/BSE) algorithm is derived to separate or extract statistically dependent source signals from their underdetermined or determined linear mixtures. Firstly, we show that maximisation NG measure can separate/extract statistically weak dependent source signals. Then, the nonparametric NG measure is defined by statistical distances between different cumulative distribution function (CDF) of separated signals, which can be estimated by quantiles and order statistics (OS) using L2 norm efficiently. Finally, the nonparametric NG measure aided algorithm is optimised by a deflation procedure. Simulation results for synthesis and real world data show that the proposed algorithm can extract the desired dependent source signals and yield expected performance.

Blind signal separation method is the method that the source signals are obtained by separating only the mixed signals. Generally, separation precision is lower when the basic method of blind signal separation is used for the sound signals. It is the cause that the power of the sound signals is concentrated in the low-frequency band. For the problem, we already considered that the method of the blind signal separation using QMF. Specifically, the mixed signals are divided into two frequency bands, and we perform suitable separation for each two bands. As the result, the conventional method has higher separation precision than the basic method. However, in the low-frequency band, power spectrum is not constant for frequency. Therefore, it is expected that the separation precision is higher than the conventional method by performing the suitable separation for each of signals based on the division to several narrow frequency bands on the low-frequency band. Then, in this paper, we propose the method that has higher separation precision and we evaluate the proposed method by computer simulation.

Under-determined blind separation becomes poorer with increasing number of tags, to the point where it cannot separate source tag signals, reducing overall system performance. This paper proposes a parallelizable identification anti-collision algorithm based on non-negative matrix factorization and adaptive ID sequence grouping of binary tree slots. The number of tags in each group can be controlled within the optimum range by selecting a reasonable number to retained source signal separation in the RFID system, which will greatly improve system performance. With the Matlab software numerical calculation and simulation, the results show that tag identification rate improves from 152.8% to 359.2% compared with the blind separation and dynamic bit-slot group algorithm using the same multi-antenna technology for 4–16 antennas, while increasing tag identification speed from 60% to 78.4%. Thus, the proposed algorithm provides high efficiency and low cost and will have very good application for fast identification of large numbers of tags.

Blind signal processing, especially independent component analysis (ICA) and multichannel blind deconvolution/equalization (MBD) problems have recently gained much interest due to many applications, especially in processing of biomedical signals (e.g. EEG, MEG, EMG, EOG, ECG), in wireless communications, 'cocktail party' problem, speech enhancement, geophysics, and source localization. In this paper both linear and nonlinear state space models for blind and semi-blind separation of linearly/nonlinearly mixed and filtered independent source signals are proposed. New unsupervised adaptive learning algorithms performing mutual independence of output signals are developed. For nonlinear mixture hyper radial basis function (HRBF), neural network proposed by Poggio and Girosi is employed and associated supervised-unsupervised learning rules for its parameters are developed.