Adversarial Network With Multiple Classifiers for Open Set Domain Adaptation

Domain adaptation aims to transfer knowledge from a domain with adequate labeled samples scarce samples. Prior research has introduced various open set settings in the literature extend applications of methods real-world scenarios. This paper focuses on type setting where target both private (‘unknown classes’) label space and shared (‘known space. However, source only ‘known classes’ Prevalent distribution-matching are inadequate such that demands smaller larger diverse more classes. For addressing this specific setting, prior introduces adversarial model uses fixed threshold for distinguishing known unknown lacks at handling negative transfers. We their propose novel multiple auxiliary classifiers. The proposed multi-classifier structure weighting module evaluates distinctive characteristics assigning weights which representative whether they likely belong classes encourage positive transfers during training simultaneously reduces gap between domains. A thorough experimental investigation shows our method outperforms existing number datasets.