Learning Efficiently- The Deep CNNs-Tree Network

In recent years, deep feature learning has been successfully applied in many fields such as visual recognition, speech recognition, and natural language processing. Based on the recent rapid development in deep learning community, applying Convolutional Neural Network (CNN) has impacted several fields. However, the number of parameters required to develop a sophisticated large CNN model becomes a problem. We aimed at this problem and presented the Deep CNNs-Tree Network model as our solution. By clustering similar channel features in the feature maps, we were able to create a tree of CNNs and replace the original CNN layer with the proposed model. Experiments on popular image datasets, the MNIST and CIFAR-10, has shown that the proposed network achieve similar performance of accuracy when compared to the traditional CNN, and only less than 5% of accuracy loss. A reduction of more than 70% parameters was observed using the proposed method.