Domain Adaptation Using Deep Learning

In last few years, deep neural networks have achieved excellent performance on different computer vision tasks. Most high-performance machine learning models are trained and tested using a large amount of annotated data which is drawn from a fixed distribution. But when the distribution of training data is different from the distribution of test data, the performance of classifier degrades. This phenomenon is known as the data shift effect. This problem can be solved by fine-tuning the deep neural network on an annotated data in target domain but collecting annotated data for every target domain is not an easy task. Domain adaptation has emerged as an alternative solution which uses the annotated data from the source domain, which is easily available, to learn a model which performs well on unlabeled target domain or scarcely labeled target domain. This paper does the empirical analysis of unsupervised domain adaptation results using digit domain adaptation datasets and ICCV domain adaptation challenge dataset. We implement adversarial domain adaptation network for solving domain adaptation problem based on a general adversarial domain adaptation framework. I have implemented adversarial domain adaptation networks using LeNet and ResNet50 architectures on the digit and ICCV VisDA domain adaptation datasets. I have achieved excellent results on different data shifts: synthetic to real and real to real, for image recognition task.