Generative Models in Machine Learning

2024/2025

Учебный год

ENG

Обучение ведется на английском языке

5

Кредиты

Кто читает:

Департамент больших данных и информационного поиска

Статус:

Курс по выбору

Когда читается:

4-й курс, 3 модуль

Преподаватель

Деркач Денис Александрович

Course Syllabus

Abstract

Deep generative models are widely used in many areas of applied machine learning. In this course, we will look at modern architectures of generative models and learning algorithms. The lectures will highlight the main approaches proposed by the beginning of 2021, and analyze their main advantages and disadvantages. The seminars will cover examples of generating images, texts, and other objects using variational autoencoders (VAE), generative adversarial networks (GANs), autoregressive models, normalizing flows, and other approaches. The assignments in the seminars are motivated by well-known applications of generative models in science and industry.

Learning Objectives

• Introduce students with modern generative models
• learning how to use variational autoencoders to generate new objects
• learning to use generative adversarial networks to generate new objects
• learning to use normalizing streams for image generation

Expected Learning Outcomes

• Explains the choice of dimension of the latent space
• Able to train autoencoders
• Understands how to train variation act-coders
• Focuses on promising developments in the field of generative models
• Can apply normalizing streams to generate images
• Uses various quality metrics to validate generative models
• Understands the difference between generative and discriminative models

Course Contents

• Introduction to Generative Models
• Generative adversarial networks and variational autoencoders
• Reversible Models
• Diffusion Models

Assessment Elements

• Homework 1
  The assignment concerns the quality assessment of generative modeling
• Homework 2
  The task is aimed at investigating the training of generative-adversarial networks.
• Homework 3
  The task is aimed at understanding the approaches to training normalizing flows.
• Homework 4
  Homework to check understanding of the concept of diffusion models.
• Project
  Implementation of the modern generative modeling algorithm according to its description in the article.
• Exam
  Examination on the theoretical foundations of generative modeling.

Interim Assessment

• 2024/2025 3rd module
  0.15 * Exam + 0.17 * Homework 1 + 0.17 * Homework 2 + 0.17 * Homework 3 + 0.17 * Homework 4 + 0.17 * Project

Bibliography

Recommended Core Bibliography

• Deep learning, Goodfellow, I., 2016
• Dhariwal, P., & Nichol, A. (2021). Diffusion Models Beat GANs on Image Synthesis.
• Глубокое обучение с точки зрения практика, Паттерсон, Дж., 2018

Recommended Additional Bibliography

• Глубокое обучение и TensorFlow для профессионалов : математический подход к построению систем искусственного интеллекта на Python, Паттанаяк, С., 2019

Authors

• DERKACH DENIS ALEKSANDROVICH