Graduation date: 1994

Thesis (M.Phil.)--City University of Hong Kong, 2006. / "Submitted to Department of Electronic Engineering in partial fulfillment of the requirements for the degree of Master of Philosophy" Includes bibliographical references (leaves 88-92)

This thesis presents a body of work in the field of reconfigurable and neuromorphic circuits and systems. Three main projects were undertaken. The first was using a Field-Programmable Analog Array (FPAA) to model the cable behavior of dendrites using analog circuits. The second was to design, lay out, and test part of a new FPAA, the RASP 2.9v. The final project was to use floating-gate programming to remove offsets in a neuromorphic FPAA, the RASP Neuron 1D.

Global climate models (GCM) have been used for nearly two decades now as a tool to investigate and analyze past, present, and future weather and climate. Even though the first several generations of climate models were very simple, today's models are very sophisticated. They use complex parameterization schemes to approximate many nonlinear physical fields. In these models, the resolution and time steps can be set to be as small or as large as desired. In either case, the model generates over 100 atmospheric variables and 20 land surface variables that can be reported daily or monthly. The Community Atmospheric Model Version Two global climate model spends over sixty percent of the time computing shortwave and longwave parameterization schemes. Our goal is to replace its shortwave scheme with empirical methods and show that accuracy of the tropospheric variables is not compromised when using these empirical methods. We found that an autoregressive moving average (ARMA) model can be used to simulate the solar radiation at the top of the model atmosphere. However, the calculated insolation value is only valid for one particular grid point. To simulate the radiation over the entire globe, many ARMA models need to be determined. We also found that large 4-10-10-1 neural networks can be used to simulate the solar radiation to within 2 W m-2. However, much smaller and manageable neural networks can be used to simulate the complete solar insolation term if the neural network only simulates the residual after the annual and diurnal cycles and removed from the field (referred to as the - method). By using the neural network in the - method and by setting the eccentricity term to a constant, we were able to cut the models processing of the solar insolation by at least a factor of four.

Sparse approximation is an important class of optimization problem in signal and image processing applications. This thesis presents an analog solution to this problem, based on the Locally Competitive Algorithm (LCA). A Hopfield-Network-like analog system, operating on sub-threshold currents is proposed as a solution. The results of the circuit components' implementation on the RASP2.8a chip, a Field Programmable Analog Array, are presented.

In the present study syngas-air diffusion flames are simulated using LES with artificial neural network (ANN) based chemical kinetics modeling and the results are compared with previous direct numerical simulation (DNS) study, which exhibits significant extinction-reignition and forms a challenging problem for ANN. The objective is to obtain speed-up in chemistry computation while still having the accuracy of stiff ODE solver. The ANN methodology is used in two ways: 1) to compute the instantaneous source term in the linear eddy mixing (LEM) subgrid combustion model used within LES framework, i.e., laminar-ANN used within LEMLES framework (LANN-LEMLES), and 2) to compute the filtered source terms directly within the LES framework, i.e., turbulent-ANN used within LES (TANN-LES), which further dicreases the computational speed. A thermo-chemical database is generated from a standalone one-dimensional LEM simulation and used to train the LANN for species source terms on grid-size of Kolmogorov scale. To train the TANN coefficients the thermo-chemical database from the standalone LEM simulation is filtered over the LES grid-size and then used for training. To evaluate the performance of the TANN methodology, the low Re test case is simulated with direct integration for chemical kinetics modeling in LEM subgrid combustion model within the LES framework (DI-LEMLES), LANN-LEMLES andTANN-LES. The TANN is generated for a low range of Ret in order to simulate the specific test case. The conditional statistics and pdfs of key scalars and the temporal evolution of the temperature and scalar dissipation rates are compared with the data extracted from DNS. Results show that the TANN-LES methodology can capture the extinction-reignition physics with reasonable accuracy compared to the DNS. Another TANN is generated for a high range of Ret expected to simulate test cases with different Re and a range of grid resolutions. The flame structure and the scalar dissipation rate statistics are analyzed to investigate success of the same TANN in simulating a range of test cases. Results show that the TANN-LES using TANN generated fora large range of Ret is capable of capturing the extinction-reignition physics with a very little loss of accuracy compared to the TANN-LES using TANN generated for the specific test case. The speed-up obtained by TANN-LES is significant compared to DI-LEMLES and LANN-LEMLES.

A connectivity adjusting learning algorithm, Optimal Brain Growth (OBG) was proposed. Contrast to the conventional training methods for the Artificial Neural Network (ANN) which focus on the weight-only optimization, the OBG method trains both weights and connectivity of a network in a single training process. The standard Back-Propagation (BP) algorithm was extended to exploit the error gradient information of the latent connection whose current weight has zero value. Based on this, the OBG algorithm makes a rational decision between a further adjustment of an existing connection weight and a creation of a new connection having zero weight. The training efficiency of a growing network is maintained by freezing stabilized connections in the further optimization process. A stabilized computational unit is also decomposed into two units and a particular set of decomposition rules guarantees a seamless local re-initialization of a training trajectory. The OBG method was tested for the multiple canonical, regression and classification problems and for a surrogate modeling of the pressure distribution on transonic airfoils. The OBG method showed an improved learning capability in computationally efficient manner compared to the conventional weight-only training using connectivity-fixed Multilayer Perceptrons (MLPs).

Thesis (Ph. D.)--University of Texas at Austin, 1999. / Vita. Includes bibliographical references (leaves 154-176). Available also in a digital version from Dissertation Abstracts.

Thesis (Ph. D.)--University of Missouri-Columbia, 1999. / Typescript. Vita. Includes bibliographical references (leaves 114-121). Also available on the Internet.

Tez (Yüksek Lisans) - Süleyman Demirel Üniversitesi, Fen Bilimleri Enstitüsü, İnşaat Mühendisliği Anabilim Dalı, 2008. / Kaynakça var.