The difference in data distributions among related, but different domains is a long standing problem for knowledge adaptation. A new method to transform the source domain knowledge to fit the target domain is proposed in this work. The proposed method uses deep learning method and limited number of samples from target domain to transform the source domain dataset. It treats the limited samples of target domain as seeds for initiating the transfer of source knowledge. Comprehensive experiments are conducted using different computational intelligence models and different datasets. Obtained results reveal that prediction models trained using the proposed method demonstrate the best performance in comparison with the same models trained with only source knowledge or deep learned features. Experiments show that models trained using proposed method have outperformed the baseline methods by at least 50% in 14 experiments out of a total of 18.