Toward Domain Adaptation for small data sets

Domain Adaptation (DA), i.e. reusing Machine Learning (ML) pre-trained models across related domains, is increasingly favored, especially in applications domains where training data is scarce. However, the adaptation process still requires data and time to adjust the model to the new domain. In this work, we propose a data selection mechanism to minimize the data required from the new domain when adapting a pre-trained ML model. Our target is to achieve short retraining time, minimize new domain data usage, and maximize transferred knowledge, lowering the overall cost of adapting pre-trained models. The result of our experiment indicates achieving a fast accuracy improvement in adapting the pre-trained model using our approach, comparing the regular one with minimum data usage.