Loading...
| Friend's email | |
| Your name | |
| Your email | |
| enter code | |
This page was sent successfuly
- Type of Document: Ph.D. Dissertation
- Language: Farsi
- Document No: 55311 (19)
- University: Sharif University of Technology
- Department: Computer Engineering
- Advisor(s): Beigy, Hamid
- Abstract:
- Recently, a new generation of machine learning tasks, namely data generation, was born by emerging deep networks and modern methods for training neural networks on one hand, and the growth of available training data for training these networks on the other hand. Although distribution estimation and sampling were well-known problems in the science of statics, deep generative models can properly generate samples from real world distributions that common statistical methods fail in them e.g., image and music generation.Due to these improvements in deep generative models, researchers have recently tried to propose deep generative models for datasets with complex structures. These structured datasets exist in a wide range of scientific fields such as social networks analysis, biology and pharmacology. Therefore, machine learning on structured data is becoming more and more important these days.In this dissertation, we study and improve deep generative models for structured data. Our main approach is to improve the performance of existing deep autoregressive generative models by injecting the available structural information of the data to them. After a quick review of the literature, we first study the problem in the case that the structure of the data is fixed i.e., all samples have the same known structure, but there exist a small number of observations. Extending an autoregressive model for non-structured data, we proposed a model that can partly compensate for the data shortage by using the available structural information. According to our empirical results, in the presence of a sufficient number of observations, the baseline autoregressive method can properly estimate the underlying density function, while its performance drastically diminishes by data shortage. But our proposed method can improve the performance of this method in the presence of a small number of observations.Next, we study the problem for variable structures, where different dimensions of the input data have unknown dependencies to each other. In fact, these structural dependencies should be implicitly learned by the model, in order to generate valid samples. For this task, also known as graph generation, we propose an autoregressive model based on transformer networks. While using the structural information of the input data, this model is capable to compute the likelihood of the input graph using just a single forward pass. We report experimental results confirming that the proposed method is competitive with the state-of-the-art models.Finally, we propose a method for density estimation in naive Bayes classification when the data is partitioned and located in multiple (distributed) sites. The goal of our proposed method is to reduce the communication load. At the end, we also suggest an idea of future work, in order to use the proposed method in probabilistic learning for structured data in distributed settings
- Keywords:
- Deep Learning ; Neural Networks ; Generative Models ; Probabilistic Graphical Models ; Graphs ; Structured Data ; Probability Density Function
Digital Object List
-
محتواي کتاب
- view
Bookmark