Benchmarking deep learning algorithms on many-core systems

• 主要作者: Oh, Jeremy Yit San
• 其他作者: He Bingsheng
• 格式: Final Year Project
• 語言: English
• 出版: 2016
• 主題: DRNTU::Engineering::Computer science and engineering
• 在線閱讀: http://hdl.handle.net/10356/66556
• 機構: Nanyang Technological University

Deep learning is a branch of machine learning that aims to extract multiple simple features from data and then combining the simple features and deriving increasingly more high level features to extract abstract data representations. In this study, we will be investigating a particular algorithm of deep learning known as a CNN (Convolutional Neural Network) for image and text classification. Specifically, this study is to note how different values for the parameters of the CNN will affect its training time and accuracy. We will also be training CNNs on different devices like the Nvidia GeForce 840M and Tesla K40 Graphic Processing Units (GPU) and seeing how they fare with different data sets. Comparisons between tensor manipulation libraries like Theano and TensorFlow will be made too, to see which is better and why. We discovered an interesting method that may make training a CNN made in Keras to be faster, and the results will be discussed in the following chapters.