Fast convolutional neural network for image classification

• Main Author: Jeon, Young Seok
• Other Authors: Chan Chok You, John
• Format: Final Year Project
• Language: English
• Published: 2017
• Subjects: DRNTU::Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence; DRNTU::Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision; DRNTU::Science::Mathematics::Applied mathematics::Optimization
• Online Access: http://hdl.handle.net/10356/72105
• Institution: Nanyang Technological University

Convolutional Neural Network(CNN) has proven its excellence in various classification tasks over the recent years. However, the major drawback of deep CNN was its days of computation time to train on large dataset with thousands of classes. The main cause for this slow computation is mostly due to the convolutional layers which performs 2D convolutions with lots of for-loops. This paper thus seeks for faster convolution algorithms such as FFT, Winograd and im2col convolutions to minimize the overall computation time and also introduces CNN written in Matlab for image classification tasks. Performance of the fast CNN algorithm written in Matlab is tested on MNIST and CIFAR-10 dataset with CNN architecture with [3*3] filter Convolutional layer sets(Convolutional, pooling and Sigmoid layers) , Fully Connected(FC) layer and followed by a Softmax layer.