A study of deep learning on many-core processors

• Main Author: Guo, Jintao
• Other Authors: He Bingsheng
• Format: Final Year Project
• Language: English
• Published: 2016
• Subjects: DRNTU::Science
• Online Access: http://hdl.handle.net/10356/66703
• Institution: Nanyang Technological University

Image object recognition in deep learning is a hot topic that many researchers have been working on. In this report, a human face recognition application was implemented using principle component analysis algorithm to study the influence of test data ratio to the accuracy of testing result. It is found that smaller test data ratio leads to better performance. Experiments have been conducted to improve the human face recognition accuracy for the feature reduction and classification process. It is observed that randomized principal component analysis with support vector machine gives the best recognition accuracy. Also deep learning techniques are applied to recognize objects in images. Various deep learning models are studied and compared for the average prediction accuracy and training time. Specifically, CaffeNet [1], VGG_CNN_M_1024 [2], and VGG16 [3] models are studied. VGG_16 performs the best when detecting objects in the image as it has the deepest neural network. The relationship between the training iteration and training accuracy has been investigated during the project. Findings have shown that training iteration and training speed are important factors for performance tuning of deep learning applications. By using this observation, the relationship between training time and training result accuracy is possible to be used to predict the training time required for a specific given accuracy.