Fig. 1

Meta-learning principles. The meta-training procedure uses a base model \(\:f\) and meta- model \(\:g\) to optimize source \(\:{S}^{\left(-t\right)}\) sample weights \(\:{w}^{}\). Parameters of the base model \(\:f\) are re-initialized and adjusted using the weighting scheme learned by the meta-model. Fine-tuning is then applied to the target domain \(\:T\).