A Novel Training Mechanism for Extending Convolutional Neural Network
Slo‐Li Chu, Cheng-Hsien Tsai · 2020
Convolutional Neural Network (CNN) has obtained great success in the computer vision domain in the recent years. These CNN models adopt deeper neural network architecture to achieve high recognition accuracy, the training costs of time and dataset are dramatically increased. While the recognizing categories are expanded, the CNN architecture needs to be modified, the whole CNN model requires to be retrained. Transfer learning method is adopted to save the training cost by migrating part of learned weights, from the existed CNN model to the target CNN model with expanded recognizing categories. However, the requirement of modifying neural network architecture still consumes huge amount of the training cost. This paper presents a new training mechanism, called Extended Learning, to solve the above problems. By using the proposed Partially Back-Propagation Operation, the CNN model can expand new classification categories without modifying the architecture of the CNN model, the learning weights from previously training results can be retained, the training cost of time and dataset can be reduced accordingly. The experimental result shows that the proposed extended learning method can save 16.7% training image count compared to the transfer learning method, with the target accuracy of 0.75.