Score-Based Generative Modeling or Diffusion Model

Please enable javascript in your browser.

Tavakoli Shiyadeh, Reza | 2024

0 Viewed

Type of Document: M.Sc. Thesis
Language: Farsi
Document No: 57519 (02)
University: Sharif University of Technology
Department: Mathematical Sciences
Advisor(s): Jafari, Amir; Rohban, Mohammad Hossein
Abstract:
In a GAN, the generator receives random noise data and produces outputs similar to real data. The generated output data, along with the original dataset, is fed to the discriminator, which distinguishes between real and fake data. The two networks have opposing objectives, but they help each other learn in practice, continuously improving each other's performance. Each of the two networks aims to minimize its own error while maximizing the other's error. John Nash mathematically proved that the optimal solution to these problems occurs when neither agent can further increase its utility. This situation is referred to as Nash equilibrium. If the training data is limited, the GAN cannot generate high-quality images, so the diffusion generator will be used. Therefore, it is necessary to address the style GAN network and GAN issues, as problems such as failure to converge, mode collapse, and vanishing gradients are significant weaknesses. Additionally, the relationships of the generative algorithm can be explained through stochastic differential equations
Keywords:
Artificial Neural Network ; Cost Function ; Generative Models ; Discriminative Model ; Generative Adversarial Networks ; Random Noise ; Diffusion Model