GAN-based Homogenous Transfer Learning Method for Regression Problems*

Buyun Wang, Shota Kato, Manabu Kano

Abstract: In industrial scenarios, the difficulty of collecting a sufficient amount of data for data analysis is a long-standing problem. Transfer learning (TL) is believed to be a potential solution to this problem. TL is a methodology that transfers useful knowledge from previous data and/or tasks in a source domain to help build a more accurate model in a target domain. So far most research is conducted in the fields of computer vision and natural language processing, and thus images or text are taken as inputs. In this study, we aim to propose a TL method for industrial scenarios; specifically, a regression task that has signal data (continuous variables) as inputs. Generative adversarial network (GAN) is widely known for its ability to generate artificial data (images, text, audio, etc) that is similar to real data. However, few previous studies have been done to investigate the performance of GAN in a regression problem setting. This study proposes a GAN-based TL method where an adjusted GAN is designed to tackle the regression problems. To evaluate the model performance of the proposed method, artificial datasets were generated to simulate different types of domain discrepancy. The experimental results suggested that the proposed method can improve the prediction accuracy in the target domain by utilizing a source domain that has similar input distributions and moderately different input-output relationships.