An efficient implementation of lattice-ladder multilayer perceptrons in field programmable gate arrays

Tomyslav Sledevič

Doctoral dissertation

Dissertations are not being sold

Quantity

Description

The implementation efficiency of electronic systems is a combination of conflicting requirements, as increasing volumes of computations, accelerating the exchange of data, at the same time increasing energy consumption forcing the researchers not only to optimize the algorithm, but also to quickly implement in a specialized hardware. Therefore in this work, the problem of efficient and straightforward implementation of operating in a real-time electronic intelligent systems on field-programmable gate array (FPGA) is tackled.

The object of research is specialized FPGA intellectual property (IP) cores that operate in a real-time. In the thesis the following main aspects of the research object are investigated: implementation criteria and techniques.

The aim of the thesis is to optimize the FPGA implementation process of selected class dynamic artificial neural networks. In order to solve stated problem and reach the goal following main tasks of the thesis are formulated: rationalize the selection of a class of Lattice-Ladder Multi-Layer Perceptron (LLMLP) and its electronic intelligent system test-bed – a speaker dependent Lithuanian speech recognizer, to be created and investigated; develop dedicated technique for implementation of LLMLP class on FPGA that is based on specialized efficiency criteria for a circuitry synthesis; develop and experimentally affirm the efficiency of optimized FPGA IP cores used in Lithuanian speech recognizer. The dissertation contains: introduction, four chapters and general conclusions. The first chapter reveals the fundamental knowledge on computer-aideddesign, artificial neural networks and speech recognition implementation on FPGA. In the second chapter the efficiency criteria and technique of LLMLP IP cores implementation are proposed in order to make multi-objective optimization of throughput, LLMLP complexity and resource utilization. The data flow graphs are applied for optimization of LLMLP computations. The optimized neuron processing element is proposed. The IP cores for features extraction and comparison are developed for Lithuanian speech recognizer and analyzed in third chapter. The fourth chapter is devoted for experimental verification of developed numerous LLMLP IP cores.

The experiments of isolated word recognition accuracy and speed for different speakers, signal to noise ratios, features extraction and accelerated comparison methods were performed. The main results of the thesis were published in 12 scientific publications: eight of them were printed in peer-reviewed scientific journals, four of them in a Thomson Reuters Web of Science database, four articles – in conference proceedings. The results were presented in 17 scientific conferences.

Read electronic version of the book:

DOI: https://doi.org/10.20334/2371-M

Book details

Data sheet

Year:
2016
ISBN:
978-609-457-933-2
Imprint No:
2371-M
Dimensions:
145×205 mm
Pages:
172 p.
Cover:
Softcover
Language:
Lithuanian
16 other books in the same category:

Follow us on Facebook