Semi-supervised adversarial discriminative domain adaptation

Abstract

Domain adaptation is a potential method to train a powerful deep neural network across various datasets. More precisely, domain adaptation methods train the model on training data and test that model on a completely separate dataset. The adversarial-based adaptation method became popular among other domain adaptation methods. Relying on the idea of GAN, the adversarial-based domain adaptation tries to minimize the distribution between the training and testing dataset based on the adversarial learning process. We observe that the semi-supervised learning approach can combine with the adversarial-based method to solve the domain adaptation problem. In this paper, we propose an improved adversarial domain adaptation method called Semi-Supervised Adversarial Discriminative Domain Adaptation (SADDA), which can outperform other prior domain adaptation methods. We also show that SADDA has a wide range of applications and illustrate the promise of our method for image classification and sentiment classification problems.

Keywords: Domain adaptation; Semi-supervised adversarial discriminative domain adaptation; Semi-supervised domain adaptation.

Fig. 1

Examples of images from different datasets. (a) Some digit images from MNIST [5], USPS [6], MNIST-M [7], and SVHN [8] datasets. (b) Some object images from the “bird” category in CALTECH [9], LABELME [10], PASCAL [11], and SUN [12] datasets

Fig. 2

An overview of the SADDA. Firstly, training the source encoder (M_s) and the classification (C_s) using the source labeled images (X_s, Y_s). Secondly, training a target encoder (M_t) through the domain adversarial process. Finally, in the testing phase, concatenate the target encoder (M_t) and the classification (C_s) to create the complete model, which will predict the label of the target dataset precisely

Fig. 3

t-SNE embedding of digit classification, using (2 x 2 x 256) dimensional representation, with Source only (on the left) and SADDA (on the right) on the target dataset. Note that SADDA minimizes intra-class distance and maximizes inter-class distance

Fig. 4

The discriminator loss and adversarial loss in the LABEL → CALTECH experiment

Fig. 5

The overview of the SADDA method for the digit recognition task. We found that Global Average Pooling (GAP) [55] increased model stability and reduce the number of parameter

Fig. 6

The overview of the SADDA method for the object recognition task on the VLCS [23] dataset. With the input image’s shape is 64 x 64 x 3

Fig. 7

The overview of the SADDA method for the sentiment classification task. The input sentence has a max length equal to 50. In the design above, to prevent overfitting, the LSTM layer is always followed by the dropout layer with 0.2 rates. The numbers under the particular layer are the output shape of that layer