Accelerating Convolutional Neural Network Systems

Convolutional Neural Networks have recently been shown to be highly effective classifiers for image and speech data. Due to the large volume of data required to build useful models, and the complexity of the models themselves, efficiency has become one of the primary concerns. This work shows that frequency domain methods can be utilised to accelerate the performance training, inference, and sliding window classification, despite the problem of CNNs using small kernels. A speedup is demonstrated on several applications including traffic sign detection and a range image classification tasks.