RODA-OOD: Robust Domain Adaptation from Out-of-Distribution Data

Abstract

Domain adaptation aims to effectively learn from two domains with different distributions, solving labeling problems; however, traditional methods assume that the source and target data are in-distribution data that share the same labels. In practice, Out-Of-Distribution (OOD) data which do not share labels with the existing data may also be collected during the target data collection process. These OOD data introduce noise and confusion, leading to decreased performance during adaptation. To address this issue, we propose RObust Domain Adaptation from Out-Of-Distribution data (RODA-OOD), a novel method based on data-centric AI principles that focuses on improving data quality rather than refining model architecture. RODA-OOD utilizes the characteristics of deep learning models that prioritize learning in-distribution data, which are easier to train on compared to OOD data. By dynamically adjusting the threshold for OOD detection, the proposed method effectively filters out OOD data, allowing the model to focus on relevant target data. RODA-OOD was compared with competitor and original domain adaptation algorithms based on target data accuracy. The results show that RODA-OOD demonstrates the most robust performance against OOD data, achieving a 21.3% increase in accuracy compared to existing domain adaptation methods. Thus, RODA-OOD can provide a solution to the OOD issue in unsupervised domain adaptation.

Keywords

• robust domain adaptation
• out-of-distribution data
• temperature scaling
• curriculum learning
• RODA-OOD