Improving Domain-Specific Classification By Collaborative Learning With Adaptation Networks

With big data becoming more and more popular,unlabeled data can be easily fetched with low cost,while labeled data does not.Labeling data manually is such a time-consuming and laborious task.How to utilize available labeled dataset to improve performance on different but related unlabeled target dataset is a task with important practical significance.Unsupervised domain adaptation aims at improving performance on target domain,with labeled source domain dataset and unlabeled target domain dataset,even if they are not totally in a same distribution.In unsupervised domain adaptation,the process of learning domain invariant feature representation is often dominated by labeled source data,while specific characteristics of target domain are usually ignored.In order to improve classification performance on target domain,we propose a new method by collaborative learning between a target specific network and an adaptation network,which is different from traditional methods that directly reduce discrepancies between source domain and target domain.To further improve the generalization ability of the target specific network,we also propose a cluster regularization method,forcing target data point moving close to accumulate cluster centers.Since the target specific network is trained only under target domain,it should gain a better classification performance on target domain,which can conversely improve the training of the adaptation network.Thus,two networks are within a collaborative relationship.We validate our proposed methods 'efficiencies and superiorities with extending experiments on multiple digit datasets and object datasets.