Train the Neural Network by Abstract Images

Like the textbook for students' learning, the training data plays a significant role in the network's training. In most cases, people intend to use big-data to train the network, which leads to two problems. Firstly, the knowledge learned by the network is out of control. Secondly, the space occupation of big-data is huge. In this paper, we use the concepts-based knowledge visualization [33] to visualize the knowledge learned by the model. Based on the observation results and information theory, we make three conjectures about the key information provided by the dataset. Finally, we use experiments to prove that the artificial abstracted data can be used in networks' training, which can solve the problem mentioned above. The experiment is designed based on Mask-RCNN, which is used to detect and classify three typical human poses on the construction site.