The Structure of Deep Neural Network for Interpretable Transfer Learning

Training a deep neural network requires a large amount of high-quality data and time. However, most of the real tasks don't have enough labeled data to train each complex model. To solve this problem, transfer learning reuses the pretrained model on a new task. However, one weakness of transfer learning is that it applies a pretrained model to a new task without understanding the output of an existing model. This may cause a lack of interpretability in training deep neural network. In this paper, we propose a technique to improve the interpretability in transfer learning tasks. We define the interpretable features and use it to train model to a new task. Thus, we will be able to explain the relationship between the source and target domain in a transfer learning task. Feature Network (FN) consists of Feature Extraction Layer and a single mapping layer that connects the features extracted from the source domain to the target domain. We examined the interpretability of the transfer learning by applying pretrained model with defined features to Korean characters classification.