Domain-Adversarial Neural Networks

• DOI: 10.48550/arxiv.1412.4446
• Publication Date: 2014-01-01
• Authors: Hana Ajakan; Pascal Germain; Hugo Larochelle; François Laviolette; Mario Marchand
• Keywords: FOS: Computer and information sciences; Computer Science - Machine Learning; Statistics - Machine Learning; 0202 electrical engineering, electronic engineering, information engineering; Computer Science - Neural and Evolutionary Computing; Machine Learning (stat.ML); 02 engineering and technology; Neural and Evolutionary Computing (cs.NE); Machine Learning (cs.LG)

We introduce a new representation learning algorithm suited to the context of domain adaptation, in which data at training and test time come from similar but different distributions. Our is directly inspired by theory on adaptation suggesting that, for effective transfer be achieved, predictions must made based that cannot discriminate between (source) (target) domains. propose objective implements this idea neural network, whose hidden layer trained predictive classification task, uninformative as input. experiments sentiment analysis benchmark, where target available unlabeled, show our network adaption has better performance than either standard or an SVM, even if input features extracted with state-of-the-art marginalized stacked denoising autoencoders Chen et al. (2012).