high performance computing on graphics processing units: hgpu.org

Deep Convolutional Neural Networks have received a lot of attention over the past few years as a promising technique for object classification in images. In this thesis, we implemented the frequency domain convolution for the popular Caffe framework. Deep Convolutional Neural Networks suffer from long training times even on contemporary hardware, which we want to address in this thesis. Particularly, the operation performed in a convolutional layer is computationally very expensive. We replaced the traditional convolution operation by a frequency domain convolution and achieved speed-ups of 2.2x and 1.8x in the forward pass and training of a well-known award-winning Deep Convolutional Neural Network, respectively. On self-designed convolutional layers, we achieved even higher speed-ups. Thus, we reduced long training and evaluation times of Deep Convolutional Neural Networks by a considerable factor. Moreover, we obtained a constant computation time for bigger kernels in a convolution layer, therefore, facilitating new structures for Deep Convolutional Neural Networks.