APPARATUS AND METHOD FOR UNSUPERVISED DOMAIN ADAPTATION

摘要

An apparatus and method for adapting an unsupervised domain are provided. An unsupervised domain adaptation apparatus according to an exemplary embodiment adapts a deep learning model in which supervised learning for a source domain is completed to an unsupervised domain adaptation to a target domain. An unsupervised domain adaptation apparatus, comprising: a plurality of first data belonging to the source domain (x _sⁱ ) and a pair of _{labels (y s}ⁱ ) for each of the first data _{(x s}ⁱ , y _sⁱ ), A forward pass is performed by inputting a plurality of second data (x _T^j ) belonging to the target domain into the deep learning model, respectively, and a Bernoulli distribution, which is a trial probability p, during the execution of the forward pass. A first learning unit for inserting a dropout following) into the deep learning model; And an uncertainty vector for each class-specific predicted value (y-prediction) and the label (y _sⁱ ) output through the forward pass, and the second data (x _T^{j) output through the forward pass.} And a second learning unit that performs back propagation to minimize uncertainty about a learning parameter of the deep learning model by making each input.