Search results for: fast convergence

Authors: Y.Ben Jemaa, M.Jaidane

Abstract:

In wireless communication system, a Decision Feedback Equalizer (DFE) to cancel the intersymbol interference (ISI) is required. In this paper, an exact convergence analysis of the (DFE) adapted by the Least Mean Square (LMS) algorithm during the training phase is derived by taking into account the finite alphabet context of data transmission. This allows us to determine the shortest training sequence that allows to reach a given Mean Square Error (MSE). With the intention of avoiding the problem of ill-convergence, the paper proposes an initialization strategy for the blind decision directed (DD) algorithm. This then yields a semi-blind DFE with high speed and good convergence.

Keywords: Adaptive Decision Feedback Equalizer, PerformanceAnalysis, Finite Alphabet Case, Ill-Convergence, Convergence speed.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 2071

Authors: Sandeep Kaur Chouhan, Jatinderdeep Kaur, S. S. Bhatia

Abstract:

The existence of sine and cosine series as a Fourier series, their L1-convergence seems to be one of the difficult question in theory of convergence of trigonometric series in L1-metric norm. In the literature so far available, various authors have studied the L1-convergence of cosine and sine trigonometric series with special coefficients. In this paper, we present a modified cosine and sine sums and criterion for L1-convergence of these modified sums is obtained. Also, a necessary and sufficient condition for the L1-convergence of the cosine and sine series is deduced as corollaries.

Keywords: Conjugate Dirichlet kernel, Dirichlet kernel, L1-convergence, modified sums.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 1220

Authors: Vishnu Pratap Singh Kirar

Abstract:

Artificial Neural Network (ANN) can be trained using back propagation (BP). It is the most widely used algorithm for supervised learning with multi-layered feed-forward networks. Efficient learning by the BP algorithm is required for many practical applications. The BP algorithm calculates the weight changes of artificial neural networks, and a common approach is to use a twoterm algorithm consisting of a learning rate (LR) and a momentum factor (MF). The major drawbacks of the two-term BP learning algorithm are the problems of local minima and slow convergence speeds, which limit the scope for real-time applications. Recently the addition of an extra term, called a proportional factor (PF), to the two-term BP algorithm was proposed. The third increases the speed of the BP algorithm. However, the PF term also reduces the convergence of the BP algorithm, and criteria for evaluating convergence are required to facilitate the application of the three terms BP algorithm. Although these two seem to be closely related, as described later, we summarize various improvements to overcome the drawbacks. Here we compare the different methods of convergence of the new three-term BP algorithm.

Keywords: Neural Network, Backpropagation, Local Minima, Fast Convergence Rate.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 3559

Authors: Sunitha. S.L., V. Udayashankara

Abstract:

Background noise is particularly damaging to speech intelligibility for people with hearing loss especially for sensorineural loss patients. Several investigations on speech intelligibility have demonstrated sensorineural loss patients need 5-15 dB higher SNR than the normal hearing subjects. This paper describes Discrete Cosine Transform Power Normalized Least Mean Square algorithm to improve the SNR and to reduce the convergence rate of the LMS for Sensory neural loss patients. Since it requires only real arithmetic, it establishes the faster convergence rate as compare to time domain LMS and also this transformation improves the eigenvalue distribution of the input autocorrelation matrix of the LMS filter. The DCT has good ortho-normal, separable, and energy compaction property. Although the DCT does not separate frequencies, it is a powerful signal decorrelator. It is a real valued function and thus can be effectively used in real-time operation. The advantages of DCT-LMS as compared to standard LMS algorithm are shown via SNR and eigenvalue ratio computations. . Exploiting the symmetry of the basis functions, the DCT transform matrix [AN] can be factored into a series of ±1 butterflies and rotation angles. This factorization results in one of the fastest DCT implementation. There are different ways to obtain factorizations. This work uses the fast factored DCT algorithm developed by Chen and company. The computer simulations results show superior convergence characteristics of the proposed algorithm by improving the SNR at least 10 dB for input SNR less than and equal to 0 dB, faster convergence speed and better time and frequency characteristics.

Keywords: Hearing Impairment, DCT Adaptive filter, Sensorineural loss patients, Convergence rate.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 2171

Authors: Eatedal A. Alabdulkreem, Hamed S. Al-Raweshidy, Maysam F. Abbod

Abstract:

Border Gateway Protocol (BGP) is the standard routing protocol between various autonomous systems (AS) in the internet. In the event of failure, a considerable delay in the BGP convergence has been shown by empirical measurements. During the convergence time the BGP will repeatedly advertise new routes to some destination and withdraw old ones until it reach a stable state. It has been found that the KEEPALIVE message timer and the HOLD time are tow parameters affecting the convergence speed. This paper aims to find the optimum value for the KEEPALIVE timer and the HOLD time that maximally reduces the convergence time without increasing the traffic. The KEEPALIVE message timer optimal value founded by this paper is 30 second instead of 60 seconds, and the optimal value for the HOLD time is 90 seconds instead of 180 seconds.

Keywords: BGP, Convergence Time, HOLD time, Keep alive.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 2041

Authors: Guiding Gu, Guo Liu

Abstract:

We discuss the convergence property of the minimum residual (MINRES) method for the solution of complex shifted Hermitian system (αI + H)x = f. Our convergence analysis shows that the method has a faster convergence than that for real shifted Hermitian system (Re(α)I + H)x = f under the condition Re(α) + λmin(H) > 0, and a larger imaginary part of the shift α has a better convergence property. Numerical experiments show such convergence properties.

Keywords: complex shifted linear system, Hermitian matrix, MINRES method.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 1625

Authors: Liang Chen

Abstract:

A family of improved secant-like method is proposed in this paper. Further, the analysis of the convergence shows that this method has super-linear convergence. Efficiency are demonstrated by numerical experiments when the choice of α is correct.

Keywords: Nonlinear equations, Secant method, Convergence order, Secant-like method.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 2046

Authors: Shu-Xin Miao

Abstract:

In this note, we discuss the convergence behavior of a modified inexact Uzawa algorithm for solving generalized saddle point problems, which is an extension of the result obtained in a recent paper [Z.H. Cao, Fast Uzawa algorithm for generalized saddle point problems, Appl. Numer. Math., 46 (2003) 157-171].

Keywords: Saddle point problem, inexact Uzawa algorithm, convergence behavior.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 1483

Authors: Giseob Byun, Ji Yeon Cho, Bong Gyou Lee

Abstract:

The purpose of this study is to analyze Green IT industry in major developed countries and to suggest overall directions for IT-Energy convergence industry. Recently, IT industry is pointed out as a problem such as environmental pollution, energy exhaustion, and high energy consumption. Therefore, Green IT gets focused which concerns as solution of these problems. However, since it is a beginning stage of this convergence area, there are only a few studies of IT-Energy convergence industry. According to this, this study examined the major developed countries in terms of institution arrangements, resources, markets and companies based on Van de Ven(1999)'s social system framework that shows relationship among key components of industrial infrastructure. Subsequently, the direction of the future study of convergence on IT and Energy industry is proposed.

Keywords: Green IT, Energy industry, Convergence, Social System Framework.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 1415

Authors: Zhenhuan Zhu, David Mulvaney, Vassilios Chouliaras

Abstract:

A new genetic algorithm, termed the 'optimum individual monogenetic genetic algorithm' (OIMGA), is presented whose properties have been deliberately designed to be well suited to hardware implementation. Specific design criteria were to ensure fast access to the individuals in the population, to keep the required silicon area for hardware implementation to a minimum and to incorporate flexibility in the structure for the targeting of a range of applications. The first two criteria are met by retaining only the current optimum individual, thereby guaranteeing a small memory requirement that can easily be stored in fast on-chip memory. Also, OIMGA can be easily reconfigured to allow the investigation of problems that normally warrant either large GA populations or individuals many genes in length. Local convergence is achieved in OIMGA by retaining elite individuals, while population diversity is ensured by continually searching for the best individuals in fresh regions of the search space. The results given in this paper demonstrate that both the performance of OIMGA and its convergence time are superior to those of a range of existing hardware GA implementations.

Keywords: Genetic algorithms, genetic hardware, machinelearning.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 2024

Authors: Jeng-Shin Sheu, Jenn-Kaie Lain, Tai-Kuo Woo, Jyh-Horng Wen

Abstract:

The least mean square (LMS) algorithmis one of the most well-known algorithms for mobile communication systems due to its implementation simplicity. However, the main limitation is its relatively slow convergence rate. In this paper, a booster using the concept of Markov chains is proposed to speed up the convergence rate of LMS algorithms. The nature of Markov chains makes it possible to exploit the past information in the updating process. Moreover, since the transition matrix has a smaller variance than that of the weight itself by the central limit theorem, the weight transition matrix converges faster than the weight itself. Accordingly, the proposed Markov-chain based booster thus has the ability to track variations in signal characteristics, and meanwhile, it can accelerate the rate of convergence for LMS algorithms. Simulation results show that the LMS algorithm can effectively increase the convergence rate and meantime further approach the Wiener solution, if the Markov-chain based booster is applied. The mean square error is also remarkably reduced, while the convergence rate is improved.

Keywords: LMS, Markov chain, convergence rate, accelerator.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 1764

Authors: Chenxue Yang, Mao Ye, Zijian Liu, Tao Li, Jiao Bao

Abstract:

Non-negative matrix factorization (NMF) is a useful computational method to find basis information of multivariate nonnegative data. A popular approach to solve the NMF problem is the multiplicative update (MU) algorithm. But, it has some defects. So the columnwisely alternating gradient (cAG) algorithm was proposed. In this paper, we analyze convergence of the cAG algorithm and show advantages over the MU algorithm. The stability of the equilibrium point is used to prove the convergence of the cAG algorithm. A classic model is used to obtain the equilibrium point and the invariant sets are constructed to guarantee the integrity of the stability. Finally, the convergence conditions of the cAG algorithm are obtained, which help reducing the evaluation time and is confirmed in the experiments. By using the same method, the MU algorithm has zero divisor and is convergent at zero has been verified. In addition, the convergence conditions of the MU algorithm at zero are similar to that of the cAG algorithm at non-zero. However, it is meaningless to discuss the convergence at zero, which is not always the result that we want for NMF. Thus, we theoretically illustrate the advantages of the cAG algorithm.

Keywords: Non-negative matrix factorizations, convergence, cAG algorithm, equilibrium point, stability.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 1697

Authors: Kankesu Jayanthakumaran, Shao-Wei Lee

Abstract:

This paper proposes the hypothesis that multilateralism and regionalism are complementary, and that regional income convergence is likely with a like minded and committed regionalism that often has links geographically and culturally. The association between international trade, income per capita, and regional income convergence in founder members of ASEAN and SAARC, is explored by applying the Lumsdaine, and Papell approach. The causal relationships between the above variables are also studied in respective trade blocs by using Granger causality tests. The conclusion is that global reforms have had a greater impact on increasing trade for both trade blocs and induced convergence only in ASEAN-5 countries. The experience of ASEAN countries shows a two-way causal relationship between the flow from trade to regional income convergence, and vice versa. There is no evidence in SAARC countries for income convergence and causality.

Keywords: ASEAN-5, SAARC-5, trade liberalisation, incomeconvergence, structural breaks and causality.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 2208

Authors: Miloje S. Radenkovic, Tamal Bose

Abstract:

This paper presents the convergence analysis of a prediction based blind equalizer for IIR channels. Predictor parameters are estimated by using the recursive least squares algorithm. It is shown that the prediction error converges almost surely (a.s.) toward a scalar multiple of the unknown input symbol sequence. It is also proved that the convergence rate of the parameter estimation error is of the same order as that in the iterated logarithm law.

Keywords: Adaptive blind equalizer, Recursive leastsquares, Adaptive Filtering, Convergence analysis.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 1454

Authors: Rado Flajs, Miran Saje

Abstract:

In the present paper the displacement-based nonconforming quadrilateral affine thin plate bending finite element ARPQ4 is presented, derived directly from non-conforming quadrilateral thin plate bending finite element RPQ4 proposed by Wanji and Cheung [19]. It is found, however, that element RPQ4 is only conditionally unisolvent. The new element is shown to be inherently unisolvent. This convenient property results in the element ARPQ4 being more robust and thus better suited for computations than its predecessor. The convergence is proved and the rate of convergence estimated. The mathematically rigorous proof of convergence presented in the paper is based on Stummel-s generalized patch test and the consideration of the element approximability condition, which are both necessary and sufficient for convergence.

Keywords: Quadrilateral thin plate bending element, convergence, generalized patch test.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 1380

Authors: Ramandeep Behl, Prashanth Maroju, S. S. Motsa

Abstract:

In this paper, we study the semilocal convergence of a fifth order iterative method using recurrence relation under the assumption that first order Fréchet derivative satisfies the Hölder condition. Also, we calculate the R-order of convergence and provide some a priori error bounds. Based on this, we give existence and uniqueness region of the solution for a nonlinear Hammerstein integral equation of the second kind.

Keywords: Hölder continuity condition, Fréchet derivative, fifth order convergence, recurrence relations.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 1935

Authors: Ahmad Alhawarat, Mustafa Mamat, Mohd Rivaie, Ismail Mohd

Abstract:

Conjugate gradient method has been enormously used to solve large scale unconstrained optimization problems due to the number of iteration, memory, CPU time, and convergence property, in this paper we find a new class of nonlinear conjugate gradient coefficient with global convergence properties proved by exact line search. The numerical results for our new βK give a good result when it compared with well known formulas.

Keywords: Conjugate gradient method, conjugate gradient coefficient, global convergence.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 2294

Authors: Samia Salah, M’hamed Hadj Sadok, Abderrezak Guessoum

Abstract:

In this paper, we propose a new modular approach called neuroglial consisting of two neural networks slow and fast which emulates a biological reality recently discovered. The implementation is based on complex multi-time scale systems; validation is performed on the model of the asynchronous machine. We applied the geometric approach based on the Gerschgorin circles for the decoupling of fast and slow variables, and the method of singular perturbations for the development of reductions models.

This new architecture allows for smaller networks with less complexity and better performance in terms of mean square error and convergence than the single network model.

Keywords: Gerschgorin’s Circles, Neuroglial Network, Multi time scales systems, Singular perturbation method.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 1605

Authors: César Garza

Abstract:

In supervised binary classification and regression problems, it is well-known that learnability is equivalent to uniform convergence of the hypothesis class, and if a problem is learnable, it is learnable by empirical risk minimization. For the general learning setting of unsupervised learning tasks, there are non-trivial learning problems where uniform convergence does not hold. We present here the task of learning centers of mass with an extra feature that “activates” some of the coordinates over the unit ball in a Hilbert space. We show that the learning problem is learnable under a stable RLM rule. We introduce a family of distributions over the domain space with some mild restrictions for which the sample complexity of uniform convergence for these problems must grow logarithmically with the dimension of the Hilbert space. If we take this dimension to infinity, we obtain a learnable problem for which the uniform convergence property fails for a vast family of distributions.

Keywords: Statistical learning theory, learnability, uniform convergence, stability, regularized loss minimization.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 354

Authors: Osama Yusuf Ababneh

Abstract:

For the last years, the variants of the Newton-s method with cubic convergence have become popular iterative methods to find approximate solutions to the roots of non-linear equations. These methods both enjoy cubic convergence at simple roots and do not require the evaluation of second order derivatives. In this paper, we present a new Newton-s method based on contra harmonic mean with cubically convergent. Numerical examples show that the new method can compete with the classical Newton's method.

Keywords: Third-order convergence, non-linear equations, root finding, iterative method.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 2964

Authors: Slimane Ouhmad, Abdellah Halimi

Abstract:

In this work, neural networks methods MLP type were applied to a database from an array of six sensors for the detection of three toxic gases. The choice of the number of hidden layers and the weight values are influential on the convergence of the learning algorithm. We proposed, in this article, a mathematical formula to determine the optimal number of hidden layers and good weight values based on the method of back propagation of errors. The results of this modeling have improved discrimination of these gases and optimized the computation time. The model presented here has proven to be an effective application for the fast identification of toxic gases.

Keywords: Back-propagation, Computing time, Fast identification, MLP neural network, Number of neurons in the hidden layer.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 2262

Authors: Sorana M. Manoiu, Razvan V. Mustata, Jiří Strouhal, Carmen G. Bonaci, Dumitru Matis, Jiřina Bokšová

Abstract:

In the current context of globalization, accountability has become a key subject of real interest for both, national and international business areas, due to the need for comparability and transparency of the economic situation, so we can speak about the harmonization and convergence of international accounting. The paper presents a qualitative research through content analysis of several reports concerning the roadmap for convergence. First, we develop a conceptual framework for the evolution of standards’ convergence and further we discuss the degree of standards harmonization and convergence between US GAAP and IAS/IFRS as to October 2012. We find that most topics did not follow the expected progress. Furthermore there are still some differences in the long-term project that are in process to be completed and other that were reassessed as a lower priority project.

Keywords: Convergence, harmonization, FASB, IASB, IFRS, US GAAP.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 3571

Authors: Seong-Mo An, Rohit Ramesh, Young-Sook Lee, Wan-Young Chung

Abstract:

The proper assessment of interaxial distance and convergence control are important factors in stereoscopic imaging technology to make an efficient 3D image. To control interaxial distance and convergence for efficient 3D shooting, horizontal 3D camera rig is designed using some hardware components like 'LM Guide', 'Goniometer' and 'Rotation Stage'. The horizontal 3D camera rig system can be properly aligned by moving the two cameras horizontally in same or opposite directions, by adjusting the camera angle and finally considering horizontal swing as well as vertical swing. In this paper, the relationship between interaxial distance and convergence angle control are discussed and intensive experiments are performed in order to demonstrate an easy and effective 3D shooting.

Keywords: Interaxial, Convergence, Stereoscopic, Horizontal 3D Camera Rig

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 2646

Authors: Zhong-xi Gao, Hou-biao Li

Abstract:

Recently, some convergent results of the generalized AOR iterative (GAOR) method for solving linear systems with strictly diagonally dominant matrices are presented in [Darvishi, M.T., Hessari, P.: On convergence of the generalized AOR method for linear systems with diagonally dominant cofficient matrices. Appl. Math. Comput. 176, 128-133 (2006)] and [Tian, G.X., Huang, T.Z., Cui, S.Y.: Convergence of generalized AOR iterative method for linear systems with strictly diagonally dominant cofficient matrices. J. Comp. Appl. Math. 213, 240-247 (2008)]. In this paper, we give the convergence of the GAOR method for linear systems with strictly doubly diagonally dominant matrix, which improves these corresponding results.

Keywords: Diagonally dominant matrix, GAOR method, Linear system, Convergence

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 1303

Authors: Zuan-De Wang, Hou-biao Li, Zhong-xi Gao

Abstract:

In this paper, the backward Ussor iterative matrix is proposed. The relationship of convergence between the backward Ussor iterative matrix and Jacobi iterative matrix is obtained, which makes the results in the corresponding references be improved and refined.Moreover,numerical examples also illustrate the effectiveness of these conclusions.

Keywords: Backward USSOR iterative matrix, Jacobi iterative matrix, convergence, spectral radius

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 1303

Authors: Guangbin Wang, Liangliang Li, Fuping Tan

Abstract:

In this paper, we give the generalized alternating twostage method in which the inner iterations are accomplished by a generalized alternating method. And we present convergence results of the method for solving nonsingular linear systems when the coefficient matrix of the linear system is a monotone matrix or an H-matrix.

Keywords: Generalized alternating two-stage method, linear system, convergence.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 1259

Authors: Sharfi Dirar, Shihab Elhaj, Ahmed El Fatih

Abstract:

Computational fluid dynamics were used to simulate and study the heated water boiler tube for both normal and rifled tube with a refinement of the mesh to check the convergence. The operation condition was taken from GARRI power station and used in a boundary condition accordingly. The result indicates the rifled tube has higher heat transfer efficiency than the normal tube.

Keywords: Boiler tube, Convergence Check, Normal Tube, Rifled Tube.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 1463

Authors: Ilona Shiller

Abstract:

This study examines regional convergence in per capita personal income in the US and Canada. We find that the disparity in real per capita income levels across US states (Canadian provinces) has declined, but income levels are not identical. Income levels become more aligned once costs of living are accounted for in relative per capita income series. US states (Canadian provinces) converge at an annual rate of between 1.3% and 2.04% (between 2.15% and 2.37%). A pattern of σ and β-convergence in per capita personal income across regions evident over the entire sample period, is reversed over 1979-1989 (1976-1990) period. The reversal may be due to sectoral or region-specific shocks that have highly persistent effects. The latter explanation might be true for half of the US and most of Canada.

Keywords: regional convergence, regional disparities, per capita income.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 1734

Authors: Hazem M. El-Bakry, Qiangfu Zhao

Abstract:

In this paper, an approach to reduce the computation steps required by fast neural networksfor the searching process is presented. The principle ofdivide and conquer strategy is applied through imagedecomposition. Each image is divided into small in sizesub-images and then each one is tested separately usinga fast neural network. The operation of fast neuralnetworks based on applying cross correlation in thefrequency domain between the input image and theweights of the hidden neurons. Compared toconventional and fast neural networks, experimentalresults show that a speed up ratio is achieved whenapplying this technique to locate human facesautomatically in cluttered scenes. Furthermore, fasterface detection is obtained by using parallel processingtechniques to test the resulting sub-images at the sametime using the same number of fast neural networks. Incontrast to using only fast neural networks, the speed upratio is increased with the size of the input image whenusing fast neural networks and image decomposition.

Keywords: Fast Neural Networks, 2D-FFT, CrossCorrelation, Image decomposition, Parallel Processing.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 2179

Authors: Zhifeng Kong

Abstract:

Over-parameterized neural networks have attracted a great deal of attention in recent deep learning theory research, as they challenge the classic perspective of over-fitting when the model has excessive parameters and have gained empirical success in various settings. While a number of theoretical works have been presented to demystify properties of such models, the convergence properties of such models are still far from being thoroughly understood. In this work, we study the convergence properties of training two-hidden-layer partially over-parameterized fully connected networks with the Rectified Linear Unit activation via gradient descent. To our knowledge, this is the first theoretical work to understand convergence properties of deep over-parameterized networks without the equally-wide-hidden-layer assumption and other unrealistic assumptions. We provide a probabilistic lower bound of the widths of hidden layers and proved linear convergence rate of gradient descent. We also conducted experiments on synthetic and real-world datasets to validate our theory.

Keywords: Over-parameterization, Rectified Linear Units (ReLU), convergence, gradient descent, neural networks.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 897