Accelerating convolutional neural networks using fine-tuned backpropagation progress

In computer vision many tasks have achieved state-of-the-art performance using convolutional neural networks (CNNs) [11], typically at the cost of massive computational complexity. A key problem of the training is the low speed of the progress. It may cost much time especially when computational resources are limited. The focus of this paper is speeding up the training progress based on fine-tuned backpropagation progress. More specifically, we train the CNNs with standard backpropagation firstly. When the feature extraction layers got better features, then we start to block the standard backpropagation in the whole layers, the loss function values only back propagates between fully connected layers. So it can not only save time but also pay more attention to train the classifier to get the same or better result compared with training with standard backpropagation all the time. Comprehensive experiments on JD (https://www.jd.com/) datasets demonstrate significant reduction in computational time, at the cost of negligible loss in accuracy.