Loading...
| Friend's email | |
| Your name | |
| Your email | |
| enter code | |
This page was sent successfuly
- Type of Document: M.Sc. Thesis
- Language: Farsi
- Document No: 52078 (19)
- University: Sharif University of Technology
- Department: Computer Engineering
- Advisor(s): Beigy, Hamid
- Abstract:
- Deep learning methods have become more popular in the past years. These methods use complex network architectures to model rich, hierarchical datasets. Although most of the research has been centered around Discriminative models, however, recently a lot of research is focused on Deep Generative Models. Two of the pioneering models in this field are Generative Adversarial Networks and Variational Auto-Encoders. In addition, knowing the structure of data helps models to search in a narrower hypothesis space. Most of the structure in datasets are models using Probabilistic Graphical Models. Using this structural information, one can achieve better parameter estimations. In the case of Generative Models, this means sampling from a more accurate model. In this thesis, we first try to use this structural information in Variational Auto Encoder. Results show that using this information improves sampling both in qualitative and quantitative evaluations.Using the proposed techniques in this thesis results were improved on MNIST dataset.The baseline model’s Jensen-Shannon divergence was 0.6909. This value was improved to 0.6229 on the proposed model with best results. Maximum mean discrepancy was also improved from 0.2302 to 0.0677 using the same proposed model.Experiments were also done on more complex data. On CIFAR dataset both Jensen-Shannon divergence and maximum mean discrepancy were improved by 0.0176 and 0.1082 respectively. These improvements were also visible in the qualitative measurements seen in the generated samples
- Keywords:
- Probabilistic Graphical Models ; Sampling ; Generating Model ; Autoencoder ; Variational Inference
Digital Object List
محتواي کتاب
Bookmark