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: 56924 (05)
- University: Sharif University of Technology
- Department: Electrical Engineering
- Advisor(s): Yassaeei Meybodi, Mohammad Hossein
- Abstract:
- Undoubtedly, machine learning and its practical applications have become essential topics in research and industry. Among these, representation learning holds special importance. The goal in this field is to find low-dimensional and meaningful representations of high-dimensional data that encapsulate essential information. Most existing methods focus only on the statistical correlations in data, which do not necessarily indicate real relationships between variables and can affect learning performance. This becomes even more problematic when the distribution of test data differs from the training data, as it makes the learned representations less generalizable. To tackle this problem, causal learning and the use of intervention data come into play. In causal representation learning, it is assumed that the hidden low-dimensional variables are structured in a causal model, and high-dimensional data are generated from them. The objective is to discover the causal graph based on observations of high-dimensional data and generalize to new and unknown distributions with the help of the causal representation obtained from the data. This thesis examines and analyzes causal representation learning methods and causal models using intervention data. Initially, basic concepts and theories in the field of causal models and independent component analysis (ICA) were reviewed. Then, through experimental design in circular linear models, the focus was on discovering causal structures and optimizing causal learning methods. The innovations introduced in this research include developing algorithms for optimizing the discovery of causal structures with the help of intervention data and presenting new approaches for causal representation learning. The results demonstrate the high potential of using intervention data to improve the accuracy and efficiency of causal learning models and representations. This research can serve as a useful guide for researchers in machine learning and meaningful data representation learning to enhance the practical applications of causal learning
- Keywords:
- Representation Learning ; Causal Structures Learning ; Independent Component Analysis (ICA) ; Experimental Design ; Structural Causal Models
Digital Object List
-
محتواي کتاب
- view
Bookmark