论文题目:Deep Multi-view Information Bottleneck
Summary
- 单个视图的信息瓶颈模型已经发展较为成熟,但尚未有在多视图上的信息瓶颈理论,本论文将信息瓶颈理论用于处理多视图数据,在学习得到单个视图的表征后,融合生成最终表征。
Problem Statement
- 现有的多视图融合的过程都是线性的,但多个视图之间的联系往往是非线性的而且是复杂的,所以需要更好的融合的过程。
Method
- 信息瓶颈:学习得到标签信息的同时去除冗余信息。
- 双视图信息瓶颈:各个视图学习得到去除冗余的表征,并使融合后的表征富含标签信息。
- 多视图信息瓶颈:
Evaluation
- 合成数据集上进行测试
- 真实数据集上进行测试
Conclusion
- 将信息瓶颈应用于多视图数据处理,增加目标信息,减少冗余信息:The model encouraged the latent representation keeping target information as much as possible while containing the information of original features as little as possible to reduce the model complexity.
- 使用变分估计的方法对互信息进行估计:Since the mutual information terms were intractable, we maximized the lower bound of the formulation instead of directly maximizing it.
- 实验测试结果:We demonstrated experiments on various synthetic and real-world datasets to show the effectiveness of the proposed method.
Notes
- 传统的多视图之间的结合方式大多是线性结合的,但多个视图之间的关系使复杂的,各个视图之间存在共享的信息,也存在互补的信息,因此多视图的表征学习和融合的过程同样重要。
- In many classification problems, the predictions can be enhanced by fusing information from different data views.
- In particular, when the information from different views complement each other, it is expected that
multi-view learning will lead to improved predictive performance.
References
- Multi-view clustering via canonical correlation analysis.
- Document clustering using word clusters via the information bottleneck method.
- Multimodal learning with deep boltzmann machines.
- A survey of multi-view machine learning.
- On deep multi-view representation learning.
- Large-margin multi-viewinformation bottleneck.