Hierarchy Training Strategy in Image Classification
Luyu Fan, Chunfang Li, Minyong Shi · 2018
With the deepening of convolutional neural networks, training a delicate network may be more difficult. This article explores which training strategies on a limited set of training data will make the network perform better, and demonstrated a terse hierarchy training strategy. We only need to do a small-scale data processing, add a priori class label for each piece based on the original dataset, the network will learn functions to classify the priori class and final class simultaneously and end to end. The knowledge learned by hierarchy A network part could guide the classification task at the hierarchy B network part. Experiments on CIFAR-10 show that network trained by this strategy will perform better and robust. Finally, the classification accuracy on the test set also has a certain improvement compared with the ordinary training strategy.