Improving deep neural networks performance using approximate arithmetic and high level synthesis

نوع: Type: thesis

مقطع: Segment: masters

عنوان: Title: Improving deep neural networks performance using approximate arithmetic and high level synthesis

ارائه دهنده: Provider: somayehkiany

اساتید راهنما: Supervisors: Dr Hatam Abdoli

اساتید مشاور: Advisory Professors: Dr Muharram Mansoorizadeh

اساتید ممتحن یا داور: Examining professors or referees: Dr Mahdi Abbasi - Dr Dezfoulian Mir hossein

زمان و تاریخ ارائه: Time and date of presentation: 2022

مکان ارائه: Place of presentation: Amphitheater

چکیده: Abstract: Nowadays, machine learning and deep learning algorithms are widely used in various fields. Numerous methods have been proposed to solve the problems of speed and execution time of these algorithms, but, it has not been reached the conclusion that these algorithms can be executed at the right time yet. Since these algorithms are used in many fields of artificial intelligence such as machine vision, speech recognition, body recognition, etc. In this study we intend to improve their efficiency as well as energy efficiency by using approximate computing. In this dissertation, a new version of approximate computational methods to reduce runtime in neural network algorithms is presented. And we have reduced the high-level synthesis approach, sacrificing accuracy as much as possible when performing calculations. On the other hand, Neural networks have fault tolerance, so the implementation of computational algorithms does not have to be very precise. As a result, we want to design the computational accuracy to the extent required and not to damage the algorithm. The result is a smaller circuit, an increase in processing speed or operating frequency, and a reduction in circuit power consumption. The purpose of this dissertation is to investigate the effective factors in improving the speed and execution time of convolutional neural network algorithms in image processing. Convolution neural network is one of the best and most widely used networks in the field of image processing and classification. The results of this study can be used not only to improve convolution efficiency, but also in many applications and machine learning algorithms. The results are calculated using Google Kolb and Vivado high-level synthesis tools. This dissertation uses the ALWANN framework and the optimization algorithm for this work is NSGA-II, as well as evolutionary and genetic processing algorithms. It has also been used in the implementation of circuits and the neural network used in this research is ResNet. The results of this study show that with 82% accuracy of classification of images, the time of calculations, maximum latency of the multiplier circuit and power consumption compared to similar work, decreased by 2.5%, 33.5% and 41.93%, respectively

فایل: ّFile: Download فایل