Transferable feature filtration network for multi-source domain adaptation

Abstract

Compared to conventional unsupervised domain adaptation, multi-source domain adaptation (MSDA) is suitable for more practical scenarios but is more challenging because the knowledge transferred to the target domain is learned from multiple source domains. Existing adversarial-based domain adaptation methods facilitate the obfuscation of domain distributions and cross-domain alignment for representations by deceiving the discriminator. However, traditional methods neglect the distinction in the transferability of the different features, resulting in untransferable domain-varying features, such as those extracted from the background, also being forced to align between domains. To this end, we propose a feature filtration mechanism and design a corresponding neural network to achieve a selective feature alignment based on the transferability of features; this is termed a transferable feature filtration network (TFFN). We construct a transferable feature learning framework containing two subprocesses, filtering, in which the attention mechanism-based filtration network automatically extracts transferable features under additional supervision, and repairing, in which we suggest enhancing the adaptability of the filtration network by repairing it with the help of an additional classifier focused on the target knowledge. Furthermore, to facilitate the matching of the source and target distributions at the class level, we propose a filtration consistency loss to enhance the cross-domain consistency of the filtering weights. Extensive experiments conducted on MSDA benchmark datasets show that the proposed method has significant advantages over the existing methods.