The paper introduces a novel approach to unsupervised domain adaptation in deep architectures, which can be trained using a large amount of labeled data from the source domain and unlabeled data from the target domain. The method promotes the emergence of "deep" features that are both discriminative for the main learning task on the source domain and invariant across domains. This is achieved by augmenting the feed-forward model with standard layers and a gradient reversal layer, which reverses the gradient during backpropagation to encourage domain-invariant features. The approach is generic and can be implemented using any deep learning package. Experimental results on various image datasets, including MNIST, SVHN, and Office datasets, demonstrate the effectiveness of the proposed method, outperforming previous state-of-the-art techniques. The method is also shown to perform well in semi-supervised settings when a small amount of labeled target data is available.The paper introduces a novel approach to unsupervised domain adaptation in deep architectures, which can be trained using a large amount of labeled data from the source domain and unlabeled data from the target domain. The method promotes the emergence of "deep" features that are both discriminative for the main learning task on the source domain and invariant across domains. This is achieved by augmenting the feed-forward model with standard layers and a gradient reversal layer, which reverses the gradient during backpropagation to encourage domain-invariant features. The approach is generic and can be implemented using any deep learning package. Experimental results on various image datasets, including MNIST, SVHN, and Office datasets, demonstrate the effectiveness of the proposed method, outperforming previous state-of-the-art techniques. The method is also shown to perform well in semi-supervised settings when a small amount of labeled target data is available.