The paper introduces a novel approach to domain adaptation in neural networks, where data from the training and test domains come from similar but different distributions. The approach is inspired by domain adaptation theory, which suggests that effective domain transfer requires predictions based on features that cannot distinguish between the training (source) and test (target) domains. The method is implemented in neural network architectures trained on labeled source data and unlabeled target data, without requiring labeled target data. As training progresses, the approach promotes the emergence of features that are both discriminative for the source domain and indiscriminate between domains. This is achieved by augmenting the neural network with standard layers and a new *gradient reversal* layer, which reverses the gradient during backpropagation. The resulting architecture can be trained using standard backpropagation and stochastic gradient descent, making it easy to implement with deep learning packages. The authors demonstrate the effectiveness of their approach on two classification problems (document sentiment analysis and image classification) and a descriptor learning task in person re-identification. They show that their method achieves state-of-the-art performance on standard benchmarks and improves over previous methods in various experiments. The approach is generic and can be applied to almost any feed-forward model, with the only non-standard component being the gradient reversal layer. The paper also provides a theoretical foundation for the approach, showing that it directly optimizes the $\mathcal{H}$-divergence between source and target distributions. The authors compare their method to other domain adaptation techniques, highlighting its advantages in terms of simplicity, effectiveness, and generalization.The paper introduces a novel approach to domain adaptation in neural networks, where data from the training and test domains come from similar but different distributions. The approach is inspired by domain adaptation theory, which suggests that effective domain transfer requires predictions based on features that cannot distinguish between the training (source) and test (target) domains. The method is implemented in neural network architectures trained on labeled source data and unlabeled target data, without requiring labeled target data. As training progresses, the approach promotes the emergence of features that are both discriminative for the source domain and indiscriminate between domains. This is achieved by augmenting the neural network with standard layers and a new *gradient reversal* layer, which reverses the gradient during backpropagation. The resulting architecture can be trained using standard backpropagation and stochastic gradient descent, making it easy to implement with deep learning packages. The authors demonstrate the effectiveness of their approach on two classification problems (document sentiment analysis and image classification) and a descriptor learning task in person re-identification. They show that their method achieves state-of-the-art performance on standard benchmarks and improves over previous methods in various experiments. The approach is generic and can be applied to almost any feed-forward model, with the only non-standard component being the gradient reversal layer. The paper also provides a theoretical foundation for the approach, showing that it directly optimizes the $\mathcal{H}$-divergence between source and target distributions. The authors compare their method to other domain adaptation techniques, highlighting its advantages in terms of simplicity, effectiveness, and generalization.