high performance computing on graphics processing units: hgpu.org

This project classifies images in Tiny ImageNet Challenge, a dataset with 200 classes and 500 training examples for each class. Three network architectures are experimented: a traditional architecture with 4 convolutional layers + 2 fully-connected layers; a Tiny GoogleNet with 3 inception layers; and a pyramid representation-based network. Tiny GoogleNet achieved the highest top-1 validation accuracy of 47%. Work is done to reduce overfitting. Dropout improves validation accuracy by 10%. Data-augmentation of random crop and horizontal flip increased validation accuracy by 10%. Rotation does not appear to improve validation accuracy. Pyramid representation shows significant computational efficiency, achieving similar top result 240% faster computation time per batch. Training accuracy converges at 65 – 70% for all three networks. Future work is to increase expressive power of network. Training was done on Torch 7 with Facebook’s Deep Learning Extension.