Multi-Source Transfer Learning Based on Inductive Knowledge-Leveraged for Medical Datasets

摘要

Transfer learning changes the limitation of the same probability distribution among domains. There are many innovative ideas of those models which are fully used the information and the knowledge from different domains. Additional knowledge by transferring learning is beneficial to improve the learning ability in target tasks. However, most multiple source domain transfer learning algorithms are developed for the specified model. The existing transfer TGHRR algorithm is suitable to one source only. Given this problem, a new multiple source transfer learning algorithm integrated with the TGHRR and the inductive knowledge of multiple domains (MS-TGHRR in brevity) is proposed. Furthermore, MS-TGHRR algorithm has been evaluated by experiments on medical datasets for classification task. Extensive experiments demonstrate the classification accuracies trained by the newly designed MS-TGHRR algorithm over the existing multiple source transfer learning algorithms.