ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2021, Vol. 58 ›› Issue (9): 1997-2012.doi: 10.7544/issn1000-1239.2021.20200569

• 人工智能 • 上一篇    下一篇

基于胶囊网络的对抗判别域适应算法

戴宏,盛立杰,苗启广   

  1. (西安电子科技大学计算机科学与技术学院 西安 710071) (西安市大数据与视觉智能关键技术重点实验室(西安电子科技大学) 西安 710071) (hdai@stu.xidian.edu.cn)
  • 出版日期: 2021-09-01
  • 基金资助: 
    国家自然科学基金项目(61772396,61902296);西安市大数据与视觉智能关键技术重点实验室项目(201805053ZD4CG37);中国博士后科学基金项目(2019M663640)

Adversarial Discriminative Domain Adaptation Algorithm with CapsNet

Dai Hong, Sheng Lijie, Miao Qiguang   

  1. (College of Computer Science and Technology, Xidian University, Xi’an 710071) (Xi’an Key Laboratory of Big Data and Intelligent Vision (Xidian University), Xi’an 710071)
  • Online: 2021-09-01
  • Supported by: 
    This work was supported by the National Natural Science Foundation of China (61772396, 61902296), the Project of Xi’an Key Laboratory of Big Data and Intelligent Vision (201805053ZD4CG37), and the China Postdoctoral Science Foundation (2019M663640).

摘要: 关于域适应算法的研究显示了对抗性学习填补源域和目标域间差异的有效性,但仍存在其局限性,即仅从2个域抽取的样本不足以保证大部分潜在空间的域不变性.注意到胶囊网络(capsule network, CapsNet)在捕获样本的表征不变性上具有较强的能力,通过结合二者得到了一种新的域适应学习算法.首先,提出了胶囊层卷积算法,并结合残差结构,使得训练更深的胶囊网络成为可能.实验表明,这种新的胶囊网络架构能够在捕获浅层特征时取得更佳的效果.其次,传统的对抗判别域适应算法使用的卷积基容易不加分辨地模糊源域与目标域的界限,进而造成判别效果的下降.因此,在VAE-GAN(variational auto-encoder, generative adversarial networks)的启发下,通过引入重建网络作为强约束,巧妙地利用了胶囊网络可调整为自编码器的特性,使得对抗判别域适应网络能够在卷积基进行迁移时,克服传统对抗判别域适应算法易发生模式崩塌的固有缺陷,保证判别器对源域与目标域内样本共性表征的敏感度.实验表明,该方法可以在不同复杂程度的域适应任务中取得较好的性能,并在关键标准数据集上取得了最先进的成果.

关键词: 胶囊网络, 对抗判别网络, 域适应, 生成对抗网络, 自编码器

Abstract: Recently, studies on domain adaptation have shown the effectiveness of adversarial learning in filling the differences between two domains, but there are still some limitations that samples taken from two domains are not enough to keep the domain invariance in potential spaces. Inspired by the fact that CapsNet(capsule network) has a strong ability to extract the invariance of features from samples, we introduce it into the domain adaptation problem. Firstly, a new convolution algorithm is devised over capsule-layer, combined with residual block, which makes it possible to train a deeper CapsNet. Results demonstrate that this new structure of CapsNet has a stronger ability to extract features. Secondly, the traditional adversarial discriminative adaptation methods have the defect that prones to blur the boundary between different domains, which in turn leads to a decline in the discriminative performance. Inspired by VAE-GAN(variational auto-encoder, generative adversarial networks), we use a reconstruction network as a strong constraint, so that the adversarial discriminative network avoids the inherent defect of mode collapse when the convolution base is transferring, and ensures that the discriminator is sensitive to the invariance of representation in different domains. Experiments on standard datasets show that our model achieves better performance in domain adaptation tasks of varying complexity.

Key words: capsule network (CapsNet), adversarial discriminative network, domain adaptation, generative adversarial networks, auto-encoder

中图分类号: