Rangkaian neural tiruan banyak digunakan dalam banyak aplikasi dan mula digunakan dalam sistem terbenam. Baru-baru ini, platform yang baru seperti Altera DE1-SOC yang mempunyai kedua-dua pemproses and FPGA telah diperkenalkan di pasaran. Apabila menggunakan platform ini, rangkaian neural tiruan boleh dilaksanakan samada di dalam pemproses menggunakan perisian atau di dalam FPGA menggunakan perlaksanaan perkakasan. Analisis perlu dijalankan untuk mengetahui samada pemproses atau FPGA lebih sesuai untuk melaksanakan rangkaian neural tiruan. Dalam projek ini, rangka kerja tentang pelaksanaan ANN pada pemproses dan FPGA dalam Altera DE1-SOC telah dibuat dan prestasi pelaksanaan tersebut tentang ketepatan, masa pelaksanaan dan penggunaan sumber telah dikaji. Beberapa model multilayer perceptron(MLP) dengan bilangan input, bilangan neuron tersembunyi dan jenis fungsi pengaktifan yang berbeza telah dilatihkan di MATLAB dulu, barulah dilaksanakan pada kedua-dua pemproses dan FPGA. Eksperimen juga telah dijalankan untuk mengukur prestasi model-model ini dalam pemproses and FPGA. Setelah membandingkan keluaran dengan ANN pada MATLAB dan mengirakan purata kuasa dua ralat (MSE), keputusan menunjukkan ANN pada pemproses mempunyai 100% ketepatan dan ANN pada FPGA mempunyai minimum 7.3 x 10-6 MSE. Manakala, ANN pada FPGA 20 kali lebih laju berbanding dengan ANN pada pemproses. Oleh itu, sekiranya satu sistem memerlukan ketepatan yang tinggi, ANN dicadang untuk dilaksanakan pada pemproses. Manakala, sekiranya masa pelaksanaan sangat penting, ANN dicadang untuk dilaksanakan pada FPGA.
_______________________________________________________________________________________________________
Artificial neural network (ANN) has been widely used in many applications and has been started to be implemented in embedded system. Recently new platform like Altera DE1-SOC that contains both processor and FPGA had been introduced. When using this type of platform, artificial neural network can be either implemented in processor using software implementation or in FPGA using hardware implementation. Analysis should be done to see whether processor or FPGA is a better choice for the ANN. In this project, framework for implementation of ANN in processor and FPGA of Altera DE1-SOC has been developed and the efficiency of implementation of ANN in processor and FPGA in terms of accuracy, execution time and resources utilization has been studied. Several multilayer perceptron (MLP) models with different number of inputs, number of hidden neurons and types of activation function have first been trained in MATLAB and after that, these trained models have been implemented in both processor and FPGA of Altera DE1-SOC. Experiments have been carried out to test and measure the performance of these MLP models in processor and FPGA. After comparing output result with ANN that run in MATLAB and computing the mean squared error (MSE), results show that the ANN in processor has 100% accuracy and ANN in FPGA has minimum MSE of 7.3 x 10-6. While ANN in FPGA is 20 times faster than ANN in processor. Therefore, if accuracy is main priority and execution time is not so important in a system, ANN is suggested to be implemented in processor. However, if execution time of ANN must be fast like less than microsecond in a system, ANN is suggested to be implemented in FPGA.
Hardware and software implementation of artificial neural network in altera de1-soc / Lim Chun Ming