Reuse of Mid-Level Feature in Deep Convolutional Neural Network

Deep convolutional neural network have led to a series of breakthroughs for computer vision task. The success of convolutional neural networks (CNNs) is attributed to their ability to learn rich image representations. However, mid-level representations (features) learned by hidden layers, which contains rich discriminative information, are not directly used for downstream tasks. In this paper, we propose a novel end-to-end model called ReuseNet, which utilizes mid-level features. More specifically, the final fully connected layer consists of the outputs of last hidden layer and also a few intermediate hidden layers. Intermediate hidden layers would receive signals directly from the loss via back propagation. ReuseNet can utilize features of different levels and also strengthen feature propagation, in order to improve the accuracy of downstream tasks. We evaluate our proposed model on two object recognition benchmark tasks: CIFAR-10 and CIFAR-100. The evaluation results show that ReuseNet significantly outperforms VGG-like networks which are trained from the bottom to up, and some other recent models.