Improving Density Estimation Using Structural Properties of Markov Random Fields

Please enable javascript in your browser.

Madani, Hatef | 2018

912 Viewed

Type of Document: M.Sc. Thesis
Language: Farsi
Document No: 50605 (19)
University: Sharif University of Technology
Department: Computer Engineering
Advisor(s): Beigy, Hamid
Abstract:
Markov Random Fields are suitable and applicable tools for modeling dependency of data dimensions; But since time complexity of parameter learning for these models is exponential with respect to data dimension size, density estimation using these models is restricted in action. In the other hands, with growing use of neural networks in many problems a class named autoregressive networks were applied for density estimation. Although the learning time of neural network parameters is not very low, there are many efforts for parameter learning acceleration. In this thesis Markov Random Fields structural properties are used in autoregressive networks. In our proposed method the hypothesis space of problem will be more restricted using local Independencies of Markov Random Field and more accurate results can be achieved using less data. For evaluating the proposed method three datasets have been used: An artificial dataset with assumed model, Asia dataset with known model and MNIST dataset with assumed model. The proposed method improved accuracy of estimation by 3% for Asia dataset, 33% for artificial dataset and 65% for MNIST dataset
Keywords:
Neural Networks ; Markov Random Field (MRF) ; Autoencoder ; Network Structural Features ; Density Estimation ; Autoregression Property