Domain Adaptation using a Gradient Reversal Layer with Instance Weighting

We propose a new method for domain adaptation that uses a gradient reversal layer (GRL) with instance weighting. Domain adaptation methods that use GRL have the latent problem of learning data that do not contribute to improving the accuracy with which the target domain is recognized. The proposed method weights each source domain sample. This enables us to control the gradients of training samples that do not contribute to improving accuracy. In an evaluation experiment using computer graphics and real image data, accuracy in recognizing the target domain improved by 5.3% compared with existing domain adaptation methods.