Domain Adaptation Method Using Activation Function with Convolution

Abstract

Image recognition using convolutional neural networks requires a large amount of training data. Since it takes a lot of effort to label a huge amount of data with teacher labels, there are many opportunities to use training data that has already been prepared. However, if the domain of the target data is different from the prepared data (e.g., different writing styles in character recognition), the performance of model trained using prepared data on the target data will be degraded. The method that aims to solve this problem is called domain adaptation. In this research, we aim to improve the performance of existing unsupervised domain adaptation methods by introducing an activation function that includes convolutional processing. In the experiments of classification tasks, we compared our method with the method of using the general ReLU activation function, and confirmed the improvement of accuracy and extracted feature distribution.