Loading...

Dual Translation Tasks Using Dual Learning

Khoshvishkaie, Ali Akbar | 2020

331 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 53034 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Beigi, Hamid
  7. Abstract:
  8. In recent years, there have been so many studies in the field of machine learning, aiming to exploit the correlation among tasks. Among those, there are some types of tasks called primal and dual, which output of one is the input of the other. Dual learning is a method in which the dual and primal tasks are trained together. Many AI tasks emerge in dual form, e.g., English to Persian translation vs. Persian to English translation and image classification vs. image generation.Recently, several methods have been proposed to utilize the correlation between dual tasks. These methods can be divided into three groups of data-level, model-level, and inference-level dual learning. They have been implemented in several applications which apart from the improvement, effectively reduced the amount of required labeled data and achieved more stable models with better generalization.In some dual tasks, there exist a stronger correlation as they are of similar functionality. Therefore, the same model architecture can be used for both, letting them transfer stronger cross-task knowledge to obtain more stable models with strong generalization. In this dissertation, we implement our methods of dual learning in two problems of image-to-image translation, and question answering and question generation. First, we will go through the previous researches in dual learning and their applications. Then we will propose two methods for this particular type of problem. Dual learning with partial sharing which inspired by multi-task learning and model-level dual learning and recurrent dual learning, inspired by the feedback mechanism in the brain's visual cortex, trying to transfer cross-task knowledge between dual tasks at the level of the model architecture. The results suggest that both methods can effectively improve the models
  9. Keywords:
  10. Deep Learning ; Duality Theory ; Dual Learning ; Dual Tasks ; Unsupervised Learning

 Digital Object List

 Bookmark

No TOC