Implement of reconfigurable stochastic artificial neural network using FPGA

Dissertant

Jasim, Manal Hammadi

University

University of Technology

Faculty

-

Department

Department of Electrical Engineering

University Country

Iraq

Degree

Ph.D.

Degree Date

2007

English Abstract

This work uses the theory of stochastic arithmetic as a solution for the FPGA implementation of a complex feed forward, multi layered neural network.

In presenting an expandable digital architecture that provides an efficient real time implementation platform for large neural networks, the architecture combines stochastic computation techniques with a novel Look Up-Table-based that fully exploits the Look Up-Table structure of many FPGAs. Basic operations of simple Stochastic ANN are mapped into a modular design.

The system control module , random pulse generating module , bit stream generating module , LFSR_32 (Liner Feedback Shift Register) sub module, modulator sub module, neuron module and bit stream converter module , are described in hardware using a schematic editor.

Thus the modules can be parameterized, providing easy scalability of the system to the different applications constraints and requirements.

On other hand, the internal implementation of the shared memory of the system is possible due to the state of the art for the chosen FPGA family using foundation LogiBLOX generator.

Thus additional speed advantage is added to the proposed design.

Finally the complete system is implemented in a single Field Programmable Gate Arrays (FPGA) platform using the Foundation 4.1i, which is a software tool from Xilinx (a vendor for the chosen FPGA platform in this work) The feasibility of the proposed ANN is demonstrated by testing it using two different applications case studies.

The objective of the first application is a Boolean Neural Network the simulation results show that the design is able to find the obtainable values for the functions of BNN, while, the objective of the second application is to find the frequency recognition for square wave with different frequencies, the simulation results show that the design is suitable for using in this field, The details of experiments are presented and the result discussed to demonstrate the potential benefits of this design platform.

Main Subjects

Information Technology and Computer Science

Topics

• Neural networks(Computer science)

• APA
• MLA
• AMA

Language

English

Data Type

Arab Theses

Record ID

BIM-305481