Journal of information and communication convergence engineering

The Korea Institute of Information and Commucation Engineering (한국정보통신학회)
권호 표지
DOI QR코드

DOI QR Code

Improving Adversarial Domain Adaptation with Mixup Regularization

  • Received : 2023.02.17
  • Accepted : 2023.05.16
  • Published : 2023.06.30
Download PDF
⟨ Previous Next ⟩

Abstract

Engineers prefer deep neural networks (DNNs) for solving computer vision problems. However, DNNs pose two major problems. First, neural networks require large amounts of well-labeled data for training. Second, the covariate shift problem is common in computer vision problems. Domain adaptation has been proposed to mitigate this problem. Recent work on adversarial-learning-based unsupervised domain adaptation (UDA) has explained transferability and enabled the model to learn robust features. Despite this advantage, current methods do not guarantee the distinguishability of the latent space unless they consider class-aware information of the target domain. Furthermore, source and target examples alone cannot efficiently extract domain-invariant features from the encoded spaces. To alleviate the problems of existing UDA methods, we propose the mixup regularization in adversarial discriminative domain adaptation (ADDA) method. We validated the effectiveness and generality of the proposed method by performing experiments under three adaptation scenarios: MNIST to USPS, SVHN to MNIST, and MNIST to MNIST-M.

Keywords

Acknowledgement

This study was supported by a Daejeon University Research Grant (2021).

References

  1. E. Tzeng, J. Hoffman, K. Saenko, and T. Darrell, "Adversarial discriminative domain adaptation," in 2017 IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, USA, pp. 2962-2971, 2017. DOI: 10.1109/CVPR.2017.316.
  2. V. Verma, A. Lamb, C. Beckham, A. Najafi, I. Mitliagkas, A. Courville, D. Lopez-Paz, and Y. Bengio, "Manifold mixup: Better representations by interpolating hidden states," in Proceedings of the 36th International Conference on Machine Learning, Long Beach, California, pp. 6438-6447, 2019. DOI: 1806.05236/arXiv.1806.05236. 1806.05236/arXiv.1806.05236
  3. H. Zhang, M. Cisse, Y. N. Dauphin, and D. Lopez-Paz, "Mixup: Beyond empirical risk minimization," in 6th International Conference on Learning Representations, Vancouver, Canada, pp. 1-13, 2018. DOI: 10.48550/arXiv.1710.09412.
  4. V. Verma, K. Kawaguchi, A. Lamb, J. Kannala, Y. Bengio, and D. Lopez-Paz, "Interpolation consistency training for semi-supervised learning," in Proceedings of the 28th International Joint Conference on Artificial Intelligence, Macao, China, pp. 3635-3641, 2019. DOI:10.24963/ijcai.2019/504.
  5. J. Hoffman, E. Tzeng, T. Park, J. Zhu, P. Isola, K. Saenko, A. Efros, and T. Darrell, "CyCADA: Cycle-consistent adversarial domain adaptation," in Proceedings of the 35th International Conference on Machine Learning, Stockholm, Sweden, pp. 1989-1998, 2018. DOI:10.48550/arXiv.1711.03213.
  6. Z. Pei, Z. Cao, M. Long, and J. Wang, "Multi-adversarial domain adaptation," in 32nd AAAI Conference on Artificial Intelligence, New Orleans, USA, pp. 3934-3941, 2018. DOI: 10.48550/arXiv.1809.02176.
  7. S. Sankaranarayanan, Y. Balaji, C. D. Castillo, and R. Chellappa, "Generate to adapt: Aligning domains using generative adversarial networks," in 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Salt Lake City, USA, pp. 8503-8512, 2018. DOI: 10.1109/CVPR.2018.00887.
  8. K. Bousmalis, N. Silberman, D. Dohan, D. Erhan, and D. Krishnan, "Unsupervised pixel-level domain adaptation with generative adversarial networks," in 2017 IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, USA, pp. 95-104, 2017. DOI: 10.1109/CVPR.2017.18.
  9. K. Saito, K. Watanabe, Y. Ushiku and T. Harada, "Maximum classifier discrepancy for unsupervised domain adaptation," in IEEE/CVF Conference on Computer Vision and Pattern Recognition, Salt Lake City, USA, pp. 3723-3732, 2018. DOI: 10.1109/CVPR.2018.00392.
  10. D. Berthelot, N. Carlini, I. Goodfellow, N. Papernot, A. Oliver and C. A. Raffel, "Mixmatch: A holistic approach to semi-supervised learning," in Advances in neural information processing systems 32 (NeurIPS 2019), Vancouver, Canada, pp. 5049-5059, 2019. DOI:10.48550/arXiv.1905.02249.
  11. J. Na and W. Hwang, "Deep learning based domain adaptation: A survey," 2022 Korean Institute of Broadcast and Media Engineers, vol. 27, no. 4, pp. 511-518, Jul. 2022. DOI: 10.5909/JBE.2022.27.4.511.
  12. K. T. Kim and J. Y. Choi, "Development of semi-supervised deep domain adaptation based face recognition using only a single training sample," Journal of Korea Multimedia Society, vol. 25, no. 10, pp. 1375-1385, Oct. 2022. DOI: 10.9717/kmms.2022.25.10.1375.
  13. L. Chen, H. Chen, Z. Wei, X. Jin, X. Tan, Y. Jin, and E. Chen, "Reusing the task-specific classifier as a discriminator: Discriminator-free adversarial domain adaptation," in 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition, New Orleans, USA, pp. 7171-7180, 2022. DOI:10.1109/CVPR52688.2022.00704.
  14. M. Chen, S. Zhao, H. Lui, and D. Cai, "Adversarial-learned loss for domain adaptation," in 34th AAAI Conference on Artificial Intelligence, New York, USA, pp. 3521-3528, 2020. DOI: 10.1609/aaai.v34i04.5757.
  15. T-D Troung, R.T.V. Chappa, X-B. Nguyen, N. Le, A.P.G. Dowling, and K. Luu, "OTAdapt: Optimal transport-based approach for unsupervised domain adaptation," in 2022 26th International Conference on Pattern Recognition, Montreal, Canada, pp. 2850-2856, 2022. DOI: 10.1109/ICPR56361.2022.9956335.