Year: 2012
Type: dissertation
Abstract: Complex-valued data arise in various applications, such as radar and array\n\tsignal processing, magnetic resonance imaging, communication systems,\n\tand processing data in the frequency domain. To deal with such data properly,\n\tneural networks are extended to the complex domain, referred to as\n\tcomplex-valued neural networks (CVNNs), allowing the network parameters\n\tto be complex numbers and the computations to follow the complex\n\talgebraic rules. Unlike the real-valued case, the nonlinear functions in the\n\tCVNNs do not have standard complex derivatives as the Cauchy-Riemann\n\tequations do not hold for them. Consequently, the traditional approach for\n\tderiving learning algorithms reformulates the problem in the real domain\n\twhich is often tedious. In this thesis, we rst develop a systematic and\n\tsimpler approach using Wirtinger calculus to derive the learning algorithms\n\tin the CVNNs. It is shown that adopting three steps: (i) computing a\n\tpair of derivatives in the conjugate coordinate system, (ii) using coordinate\n\ttransformation between real and conjugate coordinates, and (iii) organizing\n\tderivative computations through functional dependency graph greatly\n\tsimplify the derivations. To illustrate, a gradient descent and Levenberg-\n\tMarquardt algorithms are considered.\n\tAlthough a single-layered network, referred to as functional link network\n\t(FLN), has been widely used in the real domain because of its simplicity\n\tand faster processing, no such study exists in the complex domain. In\n\tthe FLN, the nonlinearity is endowed in the input layer by constructing\n\tlinearly independent basis functions in addition to the original variables.\n\tWe design a parsimonious complex-valued FLN (CFLN) using orthogonal\n\tleast squares (OLS) method, where the basis functions are multivariate\n\tpolynomial terms. It is observed that the OLS based CFLN yields simple\n\tstructure with favorable performance comparing to the multilayer CVNNs\n\tin several applications.\n\tIt is well known and interesting that a complex-valued neuron can solve\n\tseveral nonlinearly separable problems, including the XOR, parity-n, and\n\tsymmetry detection problems, which a real-valued neuron cannot. With this\n\tmotivation, we perform an empirical study of classi cation performance of\n\tsingle-layered CVNNs on several real-world benchmark classi cation problems\n\twith two new activation functions. The experimental results exhibit\n\tthat the classi cation performances of single-layered CVNNs are comparable\n\tto those of multilayer real-valued neural networks. Further enhancement\n\tof discrimination ability has been obtained using the ensemble approach.
Author: Md. Faijul Amin
Institution: University of Fukui
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Keywords: complex-valued neural networks, Wirtinger calculus, gradient descent, Levenberg-Marquardt algorithm, functional link network, orthogonal least squares, complex-valued neurons, XOR problem, parity problem, symmetry detection, activation functions, ensemble learning, classification, complex-valued data
Open Access status: green
OpenAlex ID: W2138187151
