Muhammad Moinuddin

The q-Least Mean Squares algorithm

The Least Mean Square (LMS) algorithm inherits slow convergence due to its dependency on the eigenvalue spread of the input correlation matrix. In this work, we resolve this problem by developing a novel variant of the LMS algorithms based on the q-derivative concept. The q-gradient is an extension of the classical gradient vector based on the concept of Jackson?s derivative. Here, we propose to minimize the LMS cost function by employing the concept of q-derivative instead of the convent ional derivative. Thanks to the fact that the q-derivative takes larger steps in the search direction as it evaluates the secant of the cost function rather than the tangent (as in the case of a conventional derivative), we show that the q-derivative gives faster convergence for q 1 when compared to the conventional derivative. Then, we present a thorough investigation of the convergence behavior of the proposed q-LMS algorithm and carry out different analyses to assess its performance. Consequently, new explicit closed-form expressions for the mean-square-error (MSE) behavior are derived. Simulation results are presented to corroborate our theoretical findings. HighlightsDevelopment of the q-gradient.Development of the q-LMS algorithm.Derivation of closed-form expressions for the mean-square-error for the proposed algorithm.Extensive simulation results are carried out to corroborate the theoretical findings.

Mean Weight Behavior of the NLMS Algorithm for Correlated Gaussian Inputs

This letter presents a novel approach for evaluating the mean behavior of the well known normalized least mean squares (NLMS) adaptive algorithm for a circularly correlated Gaussian input. The mean analysis of the NLMS algorithm requires the calculation of some normalized moments of the input. This is done by first expressing these moments in terms of ratios of quadratic forms of spherically symmetric random variables and finding the cumulative density function (CDF) of these variables. The CDF is then used to calculate the required moments. As a result, we obtain explicit expressions for the mean behavior of the NLMS algorithm.

A Novel Kernel for RBF Based Neural Networks

Radial basis function (RBF) is well known to provide excellent performance in function approximation and pattern classification. The conventional RBF uses basis functions which rely on distance measures such as Gaussian kernel of Euclidean distance (ED) between feature vector and neuron’s center, and so forth. In this work, we introduce a novel RBF artificial neural network (ANN) where the basis function utilizes a linear combination of ED based Gaussian kernel and a cosine kernel where the cosine kernel computes the angle between feature and center vectors. Novelty of the proposed work relies on the fact that we have shown that there may be scenarios where the two feature vectors (FV) are more prominently distinguishable via the proposed cosine measure as compared to the conventional ED measure. We discuss adaptive symbol detection for multiple phase shift keying (MPSK) signals as a practical example to show where the angle information can be pivotal which in turn justifies our proposed RBF kernel. To corroborate our theoretical developments, we investigate the performance of the proposed RBF for the problems pertaining to three different domains. Our results show that the proposed RBF outperforms the conventional RBF by a remarkable margin.

Efficient iris recognition system based on dual boundary detection using robust variable learning rate Multilayer Feed Forward neural network

This paper presents a novel approach towards iris recognition based on dual boundary (Pupil-Iris & Sclera-Iris) detection and then using a modified Multilayer Feed Forward neural network (MFNN) to perform an efficient automatic classification. The novelty of the work resides in the fact that the proposed method features the localization of the dual iris boundaries to be used as feature vector for classification. The process of information extraction starts by preprocessing the eye-image to remove specular highlight and then locating the pupil of the eye by using edge detection. The centroid of the detected pupil is chosen as the reference point for extracting the boundary points. The boundary points are recorded using radius vector functions approach. The proposed feature vector is obtained by concatenating the contour points of the Pupil-Iris boundary and the Sclera-Iris boundary which will yield a unique pattern named as Iris signature. The proposed method is translational and scale invariant. The classification is performed using the MFNN via a modified version of back-propagation algorithm which uses a time varying learning rate. The proposed system has been tested on moderate no of pictures taken from MMU iris database in the presence of additive noise for different values of signal-to-noise ratio (SNR). Experimental result for percentage recognition shows that the proposed method outperforms the single boundary method.

Performance Analysis of Beamforming in MU-MIMO Systems for Rayleigh Fading Channels

This paper characterizes the performance metrics of MU-MIMO systems under Rayleigh fading channels in the presence of both cochannel interference and additive noise with unknown channel state information and known correlation matrices. In the first task, we derive analytical expressions for the cumulative distribution function of the instantaneous signal-to-interference-plus-noise ratio (SINR) for any deterministic beamvectors. As a second task, exact closed-form expressions are derived for the instantaneous capacity, the upper bound on ergodic capacity, and the Gram-Schmidt orthogonalization-based ergodic capacity for similar intra-cell correlation coefficients. Finally, we present the utility of several structured-diagonalization techniques, which can achieve the tractability for the approximate solution of ergodic capacity for both similar as well as different intra-cell correlation matrices. The novelty of this paper is to formulate the received SINR in terms of indefinite quadratic forms, which allows us to use complex residue theory to characterize the system behavior. The analytical expressions obtained closely match simulation results.

Image Retrieval Based on Color and Texture Feature Using Artificial Neural Network

Content-based image retrieval CBIR is a technique that helps in searching a user desired information from a huge set of image files and interpret user intentions for the desired information. The retrieval of information is based on features of image like colour, shape, texture, annotation etc. Many of the existing methods focus on the feature extraction and to bridge up the gap between low level features and high level semantics. In this paper we propose a supervised machine learning (SML) using artificial neural network (ANN) and singular value decomposition (SVD) for image retrieval. Specifically we use back propagation algorithm (multilayer perceptron) (MLP) for training and testing our proposed model. Experimental results show that by changing parameters of feature vector back propagation method can have 62% precision instead of 49% as claimed by in Hyoung Ku LEE, Suk In Yoo [1].

Family of state space least mean power of two-based algorithms

In this work, a novel family of state space adaptive algorithms is introduced. The proposed family of algorithms is derived based on stochastic gradient approach with a generalized least mean cost function J[k]=E[∥ε[k]∥2L] for any integer L. Since this generalized cost function is having power `2L’, it includes the whole family of the power of two-based algorithms by having different values of L. The novelty of the work resides in the fact that such a cost function has never been used in the framework of state space model. It is a well-known fact that the knowledge of state space model improves the estimation of state parameters of that system. Hence, by employing the state space model with a generalized cost function, we provide an efficient way to estimate the state parameters. The proposed family of algorithms inherit simplicity in its structure due to the use of stochastic gradient approach in contrast to the other model-based algorithms such as Kalman filter and its variants. This fact is supported by providing a comparison of the computational complexities of these algorithms. More specifically, the proposed family of algorithms has computational complexity far lesser than that of the Kalman filter. The stability of the proposed family of algorithms is analysed by providing the convergence analysis. Extensive simulations are presented to provide concrete justification and to compare the performances of the proposed family of algorithms with that of the Kalman filter.

A simple approach to evaluate the ergodic capacity and outage probability of correlated Rayleigh diversity channels with unequal signal-to-noise ratios

In this article, we propose a novel method to derive exact closed-form ergodic capacity and outage probability expressions for correlated Rayleigh fading channels with receive diversity. Unlike the existing works, the proposed method employ a simple approach for the capacity and outage analysis for receiver diversity channels operating at different signal-to-noise ratios depicted in the diagonal elements of matrix Ω. With x being the channel gain vector, random variable of the form Y(a)=a + x∗Ωx is considered. Novelty of the work resides in the fact that the distribution of Y(a) is accurately determined by employing Fourier representation of unit step function followed by complex integration in a straight forward way. The ergodic channel capacity is thus calculated by using the first-order moment, E[log2(Y(1))], while the outage probability for a certain threshold γ0is evaluated using ∫0γ0fY(0)(y)dy. Extensive experiments have been conducted demonstrating the accuracy of the proposed approach.

Design of MAI constrained decision feedback equalizer for MIMO CDMA system

A decision feedback equalizer (DFE) utilizes the previous detectors assessment to get rid of the inter-symbol interference (ISI) on the received symbols. It is well analyzed in the literature that the DFE performs better than the linear equalizer when the ISI is sever because of its inherent nonlinear nature. Equalization of wireless multiple-input multiple-output (MIMO) frequency-selective channels is a challenging task mainly due to the fact that the respective MIMO equalizers should cope with inter-symbol, as well as inter-stream interference. Different techniques have been proposed for MIMO DFE. In this paper we have developed a constrained MIMO DFE for CDMA system based on constrained optimization. This is achieved by minimizing the conventional mean-square-error criterion subject to the variance of the multiple-access interference (MAI) plus noise. The novelty of the work resides in the fact that such a constrained optimization has never been employed in the design of MIMO DFE. The proposed MIMO DFE algorithm is tested in different fading environments and its performance is compared with that of the conventional least mean-square (LMS) and normalized LMS (NLMS) algorithms. Simulation results show that the proposed constrained MIMO DFE outperforms the conventional MIMO DFEs based on LMS and NLMS algorithms.

A Novel Fractional Gradient-Based Learning Algorithm for Recurrent Neural Networks

In this research, we propose a novel algorithm for learning of the recurrent neural networks called as the fractional back-propagation through time (FBPTT). Considering the potential of the fractional calculus, we propose to use the fractional calculus-based gradient descent method to derive the FBPTT algorithm. The proposed FBPTT method is shown to outperform the conventional back-propagation through time algorithm on three major problems of estimation namely nonlinear system identification, pattern classification and Mackey–Glass chaotic time series prediction.