Parameters Sharing in Residual Neural Networks

Deep neural networks (DNN) have achieved great success in machine learning due to their powerful ability to learn and present knowledge. However, models of such DNN often have massive trainable parameters, which lead to big resource burden in practice. As a result, reducing the amount of parameters and preserving its competitive performance are always critical tasks in the field of DNN. In this paper, we focused on one type of convolution neural network that has many repeated or same-structure convolutional layers. Residual net and its variants are widely used, making the deeper model easy to train. One type block of such a model contains two convolutional layers, and each block commonly has two trainable parameter layers. However, we used only one layer of trainable parameters in the block, which means that the two convolutional layers in one block shared one layer of trainable parameters. We performed extensive experiments for different architectures of the Residual Net with trainable parameter sharing on the CIFAR-10, CIFAR-100, and ImageNet datasets. We found that the model with trainable parameter sharing can obtain fewer errors on the training datasets and had a very close recognition accuracy (within 0.5%), compared to the original models. The parameters of the new model were reduced by more than 1/3 of the total of the original.