Self training domain adaptation for real world application of deep learning model

We propose a self-training domain adaptation model for real world applications of machine learning model. Our suggested model learns representation function that embedding of source and target data with same label has similar distribution. We have experimented three domain adaptation environments: grey scale image to RGB image, application to different image size, and application to real world data. Our model showed compelling performance for classification task. We also showed effectiveness of our proposed method by visualizing hidden representation of our model. We expect that our method can broaden the application of machine learning in the real world like detecting defects on new product by using the data of existing product.