Designing a Partially Understandable Neural Network through Semantic Embedding

Chengfu Tang, Dawei Dai, Haoyue Bai, Guoyin Wang, Shuyin Xia, Feng Hu

Abstract: In recent years, the convolution neural network (CNN) has been successfully applied in numerous fields as a machine learning model. However, such neural models are still considered to be “black box” for most tasks. The fundamental issue underlying this problem is that the information and knowledge learned by a neural network during the training process are unknown and unpredictable. In this study, we attempted to design a partially understandable neural network through semantic embedding. Firstly, we selected several understandable feature extraction operators as expected information. Secondly, we embedded these operators into the hierarchical layers of a neural network at the beginning of its training process. Finally, these embedded operators were only involved in forward calculation and remained unchanged in the error back-propagation of the training process. We applied our method to the ResNet and DenseNet models in image classification tasks. In the experiments, our new models achieved almost the same performance as the original ones, but decreased performance was exhibited when shutting the embedded parts, and the features in the embedded parts were understandable. The experiments verified that the embedded parts not only contributed to the classification tasks, but also caused the neural network model to be partially understandable.