AI帮你理解科学

AI 生成解读视频

AI抽取解析论文重点内容自动生成视频


pub
生成解读视频

AI 溯源

AI解析本论文相关学术脉络


Master Reading Tree
生成 溯源树

AI 精读

AI抽取本论文的概要总结


微博一下
By learning a meta-representation specified for evolving domain adaptation, we are able to capture and harness the smooth evolvement of the target domain in knowledge transfer

Learning to Adapt to Evolving Domains

NIPS 2020, (2020)

被引用1|浏览56
EI
下载 PDF 全文
引用
微博一下

摘要

Domain adaptation aims at knowledge transfer from a labeled source domain to an unlabeled target domain. Current domain adaptation methods have made substantial advances in adapting discrete domains. However, this can be unrealistic in real-world applications, where target data usually come in an online and continually evolving manner, po...更多

代码

数据

0
简介
  • Many machine learning applications require consistent performance across datasets of different underlying data distributions.
  • Following the terms of meta-learning [4], the authors formulate this setting as evolving domain adaptation (EDA): (1) the authors have access to adequate labeled examples from the source domain, and part of the target unlabeled data from a target domain evolving over time in the meta-training phase, (2) new target data of the meta-testing phase arrive sequentially online from the same evolving target distribution and cannot be stored, and (3) after adapting to these new target data sequentially, the authors test the model on all target data of meta-testing.
  • An example of the EDA problem is provided in Figure 1
重点内容
  • Many machine learning applications require consistent performance across datasets of different underlying data distributions
  • Following the terms of meta-learning [4], we formulate this setting as evolving domain adaptation (EDA): (1) we have access to adequate labeled examples from the source domain, and part of the target unlabeled data from a target domain evolving over time in the meta-training phase, (2) new target data of the meta-testing phase arrive sequentially online from the same evolving target distribution and cannot be stored, and (3) after adapting to these new target data sequentially, we test the model on all target data of meta-testing
  • Replay methods [28, 33, 38, 30] need storing samples from previous tasks or using large generative models [10], which violate the setting of EDA and are unrealistic with the limited resources of online devices such as aforementioned self-driving agents. Aiming to tackle both challenges, this paper proposes Evolution Adaptive Meta-Learning (EAML), a meta-adaptation framework to adapt to continually evolving target domain without forgetting
  • Classic domain adaptation methods are not suitable for the evolving target domains
  • We propose a meta-adaptation framework to solve evolving domain adaptation (EDA) efficiently
  • By learning a meta-representation specified for EDA, we are able to capture and harness the smooth evolvement of the target domain in knowledge transfer
方法
  • The authors evaluate the method with evolving domain datasets in different scenarios.
  • Details on datasets and implementation are deferred to appendix2.
  • This dataset consists of MNIST digits of various rotations.
  • This is similar to the protocol of [3], but they only consider discrete rotations and each domain has 60000 training samples
结果
  • Results on Rotated MNIST are provided in Table 1.
  • EAML with only meta-representation (EAML-rep) improves performance on all target data and especially the last-adapted target data, since it learns representations specified for evolving target which harness knowledge on the previous target data.
  • In Figure 3(b), the authors show the change of accuracy on the first target domain during the adaptation to the following domains.
  • Note that as EAML-rep adapts to the following target data, the accuracy on the previous target first rises, validating that the meta-representations harness knowledge from previous target when adapting to the current target.
  • EAML-full further incorporates the meta-adapter to overcome catastrophic forgetting
结论
  • The authors address the problem of adapting to evolving domains.
  • The authors propose a meta-adaptation framework to solve evolving domain adaptation (EDA) efficiently.
  • By learning a meta-representation specified for EDA, the authors are able to capture and harness the smooth evolvement of the target domain in knowledge transfer.
  • This paper opens up future questions for evolving domain adaptation.
  • How can the authors capture the intrinsic structure of evolving data more efficiently?
  • Could the authors extent the EDA framework to heterogeneous transfer learning?
  • How can the authors capture the intrinsic structure of evolving data more efficiently? Could the authors extent the EDA framework to heterogeneous transfer learning? the authors hope the work inspires further studies to pursue real-world domain adaptation applications
总结
  • Introduction:

    Many machine learning applications require consistent performance across datasets of different underlying data distributions.
  • Following the terms of meta-learning [4], the authors formulate this setting as evolving domain adaptation (EDA): (1) the authors have access to adequate labeled examples from the source domain, and part of the target unlabeled data from a target domain evolving over time in the meta-training phase, (2) new target data of the meta-testing phase arrive sequentially online from the same evolving target distribution and cannot be stored, and (3) after adapting to these new target data sequentially, the authors test the model on all target data of meta-testing.
  • An example of the EDA problem is provided in Figure 1
  • Methods:

    The authors evaluate the method with evolving domain datasets in different scenarios.
  • Details on datasets and implementation are deferred to appendix2.
  • This dataset consists of MNIST digits of various rotations.
  • This is similar to the protocol of [3], but they only consider discrete rotations and each domain has 60000 training samples
  • Results:

    Results on Rotated MNIST are provided in Table 1.
  • EAML with only meta-representation (EAML-rep) improves performance on all target data and especially the last-adapted target data, since it learns representations specified for evolving target which harness knowledge on the previous target data.
  • In Figure 3(b), the authors show the change of accuracy on the first target domain during the adaptation to the following domains.
  • Note that as EAML-rep adapts to the following target data, the accuracy on the previous target first rises, validating that the meta-representations harness knowledge from previous target when adapting to the current target.
  • EAML-full further incorporates the meta-adapter to overcome catastrophic forgetting
  • Conclusion:

    The authors address the problem of adapting to evolving domains.
  • The authors propose a meta-adaptation framework to solve evolving domain adaptation (EDA) efficiently.
  • By learning a meta-representation specified for EDA, the authors are able to capture and harness the smooth evolvement of the target domain in knowledge transfer.
  • This paper opens up future questions for evolving domain adaptation.
  • How can the authors capture the intrinsic structure of evolving data more efficiently?
  • Could the authors extent the EDA framework to heterogeneous transfer learning?
  • How can the authors capture the intrinsic structure of evolving data more efficiently? Could the authors extent the EDA framework to heterogeneous transfer learning? the authors hope the work inspires further studies to pursue real-world domain adaptation applications
表格
  • Table1: Classification Accuracy (%) on rotated MNIST dataset
  • Table2: Comparison of different settings of continual of evolving domain adaptation
Download tables as Excel
相关工作
  • Classic Domain Adaptation. Classic domain adaptation learns a representation where the domain discrepancy is minimized. [24, 7, 8] map the source and target domains into a new feature space. [19, 36] incorporate the maximum mean discrepancy (MMD). DANN [6] trains a domain discriminator to distinguish discrete source and target while the features are learned adversarially to confuse the discriminator. [35, 20, 34] further improve DANN and achieve significant performance gain. MCD [31] uses an alternative approach: maximizing the disagreement of two classifiers on the target domain. [18, 32, 12] enable pixel-level adaptation with generative architectures. These methods are tailored to discrete source and target domains and cannot be applied to EDA directly.

    Continuous Domain Adaptation. The problem of continuous domain adaptation is related to EDA but with a different context and focus. [3, 15] learn discrete target tasks online, but they do not address the evolving nature of target domains to harness it in knowledge transfer. [11, 2] focus on continually shifting target domains, but they test the model online and did not address the catastrophic forgetting. In this paper, we tackle a more challenging and realistic EDA scenario, addressing both challenges of target domain evolvement and catastrophic forgetting. A comparison of several different settings of continual or evolving domain adaptation is provided in Table 2.
基金
  • Acknowledgments and Disclosure of Funding This work was supported by the Natural Science Foundation of China (61772299, 71690231), Beijing Nova Program (Z201100006820041), and University S&T Innovation Plan by the Ministry of Education of China
研究对象与分析
training samples: 60000
Details on datasets and implementation are deferred to appendix2.

4.1 Datasets Rotated MNIST

This dataset consists of MNIST digits of various rotations. This is similar to the protocol of [3], but they only consider discrete rotations and each domain has 60000 training samples.

training samples: 60000
4.1 Datasets Rotated MNIST: This dataset consists of MNIST digits of various rotations. This is similar to the protocol of [3], but they only consider discrete rotations and each domain has 60000 training samples. 2Codes are available at https://github.com/Liuhong99/EAML

samples: 100
Images with rotation 0◦ belong to the labeled source domain. For each rotation, we have access to 100 samples of images. In meta-training, we use rotation 0−60◦

training samples: 100
In meta-testing, we test the model’s performance online with trajectory T = {X120◦ , X126◦ · · · X174◦ }. Note that in meta-testing, we have only 100 training samples for each rotation, making this task very challenging. Evolving Vehicles: This dataset contains sedans and trucks from the 1970s to 2010s (See Figure 1), which involves more complex continuous domain shift compared to rotated MNIST

引用论文
  • S. Ben-David, J. Blitzer, K. Crammer, A. Kulesza, F. Pereira, and J. W. Vaughan. A theory of learning from different domains. Machine Learning, 79(1-2):151–175, 2010.
    Google ScholarLocate open access versionFindings
  • A. Bitarafan, M. S. Baghshah, and M. Gheisari. Incremental evolving domain adaptation. IEEE Transactions on Knowledge and Data Engineering, 28(8):2128–2141, 2016.
    Google ScholarLocate open access versionFindings
  • A. Bobu, E. Tzeng, J. Hoffman, and T. Darrell. Adapting to continuously shifting domains. In International Conference on Learning Representations (ICLR) Workshop Track, 2018.
    Google ScholarLocate open access versionFindings
  • C. Finn, P. Abbeel, and S. Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 1126–1135. PMLR, 2017.
    Google ScholarLocate open access versionFindings
  • L. Franceschi, M. Donini, P. Frasconi, and M. Pontil. Forward and reverse gradient-based hyperparameter optimization. In Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 1165–1173. PMLR, 2017.
    Google ScholarLocate open access versionFindings
  • Y. Ganin, E. Ustinova, H. Ajakan, P. Germain, H. Larochelle, M. Marchand, and V. Lempitsky. Domain-adversarial training of neural networks. Journal of Machine Learning Research, 17(1):2096–2030, 2016.
    Google ScholarLocate open access versionFindings
  • B. Gong, K. Grauman, and F. Sha. Connecting the dots with landmarks: Discriminatively learning domain-invariant features for unsupervised domain adaptation. In ICML, 2013.
    Google ScholarFindings
  • B. Gong, Y. Shi, F. Sha, and K. Grauman. Geodesic flow kernel for unsupervised domain adaptation. In CVPR, 2012.
    Google ScholarLocate open access versionFindings
  • R. Gong, W. Li, Y. Chen, and L. V. Gool. Dlow: Domain flow for adaptation and generalization. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019.
    Google ScholarLocate open access versionFindings
  • I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems 27, pages 2672–2680, 2014.
    Google ScholarLocate open access versionFindings
  • J. Hoffman, T. Darrell, and K. Saenko. Continuous manifold based adaptation for evolving visual domains. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2014.
    Google ScholarLocate open access versionFindings
  • J. Hoffman, E. Tzeng, T. Park, J. Zhu, P. Isola, K. Saenko, A. A. Efros, and T. Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In Proceedings of the 35th International Conference on Machine Learning (ICML), pages 1994–2003, 2018.
    Google ScholarLocate open access versionFindings
  • Y. Jang, H. Lee, S. J. Hwang, and J. Shin. Learning what and where to transfer. In Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 3030–3039. PMLR, 2019.
    Google ScholarLocate open access versionFindings
  • J. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. Veness, G. Desjardins, A. A. Rusu, K. Milan, J. Quan, T. Ramalho, and A. Grabska-Barwinska. Overcoming catastrophic forgetting in neural networks. Proceedings of National Academy of Science USA, 114(13):3521–3526, 2017.
    Google ScholarLocate open access versionFindings
  • Q. Lao, X. Jiang, M. Havaei, and Y. Bengio. Continuous domain adaptation with variational domain-agnostic feature replay. 2020.
    Google ScholarFindings
  • Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998.
    Google ScholarLocate open access versionFindings
  • S.-W. Lee, J.-H. Kim, J. Jun, J.-W. Ha, and B.-T. Zhang. Overcoming catastrophic forgetting by incremental moment matching. In Advances in Neural Information Processing Systems 30, pages 4652–4662. 2017.
    Google ScholarLocate open access versionFindings
  • M.-Y. Liu, T. Breuel, and J. Kautz. Unsupervised image-to-image translation networks. In I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems 30, pages 700–708. Curran Associates, Inc., 2017.
    Google ScholarLocate open access versionFindings
  • M. Long, Y. Cao, J. Wang, and M. I. Jordan. Learning transferable features with deep adaptation networks. In Proceedings of the 32nd International Conference on Machine Learning (ICML), pages 97–105, 2015.
    Google ScholarLocate open access versionFindings
  • M. Long, Z. Cao, J. Wang, and M. I. Jordan. Conditional adversarial domain adaptation. In Advances in Neural Information Processing Systems 31, pages 1640–1650. 2018.
    Google ScholarLocate open access versionFindings
  • M. Long, H. Zhu, J. Wang, and M. I. Jordan. Deep transfer learning with joint adaptation networks. In Proceedings of the 34th International Conference on Machine Learning (ICML), pages 2208–2217, 2017.
    Google ScholarLocate open access versionFindings
  • Y. Mansour, M. Mohri, and A. Rostamizadeh. Domain adaptation: Learning bounds and algorithms. In The 22nd Conference on Learning Theory, Montreal, Quebec, Canada, June 18-21, 2009, 2009.
    Google ScholarLocate open access versionFindings
  • M. Mccloskey and N. J. Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. Psychology of Learning and Motivation, 24:109–165, 1989.
    Google ScholarLocate open access versionFindings
  • S. J. Pan, I. W. Tsang, J. T. Kwok, and Q. Yang. Domain adaptation via transfer component analysis. TNNLS, 22(2):199–210, 2011.
    Google ScholarLocate open access versionFindings
  • S. J. Pan and Q. Yang. A survey on transfer learning. IEEE Transactions on Knowledge and Data Engineering, 22(10):1345–1359, 2010.
    Google ScholarLocate open access versionFindings
  • J. Quionero-Candela, M. Sugiyama, A. Schwaighofer, and N. D. Lawrence. Dataset Shift in Machine Learning. The MIT Press, 2009.
    Google ScholarFindings
  • A. Raghu, M. Raghu, S. Bengio, and O. Vinyals. Rapid learning or feature reuse? towards understanding the effectiveness of maml. In International Conference on Learning Representations, 2019.
    Google ScholarLocate open access versionFindings
  • S.-A. Rebuffi, A. Kolesnikov, G. Sperl, and C. H. Lampert. icarl: Incremental classifier and representation learning. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017.
    Google ScholarLocate open access versionFindings
  • M. Riemer, I. Cases, R. Ajemian, M. Liu, I. Rish, Y. Tu, and G. Tesauro. Learning to learn without forgetting by maximizing transfer and minimizing interference. In International Conference on Learning Representations, 2018.
    Google ScholarLocate open access versionFindings
  • D. Rolnick, A. Ahuja, J. Schwarz, T. Lillicrap, and G. Wayne. Experience replay for continual learning. In Advances in Neural Information Processing Systems 32, pages 350–360. 2019.
    Google ScholarLocate open access versionFindings
  • K. Saito, K. Watanabe, Y. Ushiku, and T. Harada. Maximum classifier discrepancy for unsupervised domain adaptation. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3723–3732, 2018.
    Google ScholarLocate open access versionFindings
  • S. Sankaranarayanan, Y. Balaji, C. D. Castillo, and R. Chellappa. Generate to adapt: Aligning domains using generative adversarial networks. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 8503–8512, 2018.
    Google ScholarLocate open access versionFindings
  • H. Shin, J. K. Lee, J. Kim, and J. Kim. Continual learning with deep generative replay. In Advances in Neural Information Processing Systems 30, pages 2990–2999. 2017.
    Google ScholarLocate open access versionFindings
  • R. Shu, H. Bui, H. Narui, and S. Ermon. A DIRT-t approach to unsupervised domain adaptation. In Proceedings of International Conference on Learning Representations, 2018.
    Google ScholarLocate open access versionFindings
  • E. Tzeng, J. Hoffman, K. Saenko, and T. Darrell. Adversarial discriminative domain adaptation. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 7167– 7176, 2017.
    Google ScholarLocate open access versionFindings
  • E. Tzeng, J. Hoffman, N. Zhang, K. Saenko, and T. Darrell. Deep domain confusion: Maximizing for domain invariance. arXiv, abs/1412.3474, 2014.
    Findings
  • L. J. P. van der Maaten and G. E. Hinton. Visualizing data using t-sne. In Journal of Machine Learning Research, volume 9, pages 2579–2605, 2008.
    Google ScholarLocate open access versionFindings
  • C. Wu, L. Herranz, X. Liu, y. wang, J. van de Weijer, and B. Raducanu. Memory replay gans: Learning to generate new categories without forgetting. In Advances in Neural Information Processing Systems 31, pages 5962–5972. 2018.
    Google ScholarLocate open access versionFindings
  • F. Zenke, B. Poole, and S. Ganguli. Continual learning through synaptic intelligence. In Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 3987–3995, 2017.
    Google ScholarLocate open access versionFindings
  • Y. Zhang, T. Liu, M. Long, and M. Jordan. Bridging theory and algorithm for domain adaptation. In Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 7404–7413, 2019.
    Google ScholarLocate open access versionFindings
  • J.-Y. Zhu, T. Park, P. Isola, and A. A. Efros. Unpaired image-to-image translation using cycleconsistent adversarial networks. In The IEEE International Conference on Computer Vision (ICCV), pages 2223–2232, 2017.
    Google ScholarLocate open access versionFindings
您的评分 :
0

 

标签
评论
小科