Gig: a Knowledge-Transferable-oriented Framework for Cross-Domain Recognition

Domain Adaptation (DA) commonly finds a shared subspace, in which the discrepancy between the source and target domains is reduced and the target samples could be correctly classified. Existing studies mainly learn domain-invariant features via one shared subspaces or by two decoupled subspaces. However, since the learning of the source and target domains may interact with each other, they may neglect that (a) the domain-specific features are unique to each domain and they are not compatible in a shared subspace; and (b) insufficient transferable features between the two domains may lead to unsatisfactory performance. To address these problems, this study introduces a three-step optimization learning framework called Guidance, Imitation, and Generalization Subspace Learning (GIG). GIG first decouples the synchronous learning of the source and target domains into three subspaces, including guidance subspace for the source domain, imitation subspace for the shared domain, and generalization subspace for the target domain, so that both the domain-invariant and domain-specific knowledge can be learned as much as possible. It then learns domain-specific features by employing spectral clustering to the Guidance and Generalization subspaces, respectively, and captures the domain-invariant knowledge by aligning the marginal distribution on the Imitation subspace. In this way, the negative impacts caused by the interactions between the source and target domains are alleviated. At last, Distilled Label Regression (DLR) is proposed to incorporate the posterior probabilities of classifiers and labels as a new semantic embedding and regress the data into the semantic embedding, so that the discriminability of the Guidance subspace is improved. Two relaxed variables are introduced to optimization, such that the range of the candidate transferable information is extended and the acquisition of extreme values is ensured. By sequentially learning these three subspaces, GIG extracts more knowledge-transferable features and achieves significant performance improvements. Experiments conducted on eight benchmark datasets demonstrate the superiority of GIG.