Lecture

Diffusion: Data Denoising and Generative Modeling

Graph Chat

Chat with Graph Search

Ask any question about EPFL courses, lectures, exercises, research, news, etc. or try the example questions below.

DISCLAIMER: The Graph Chatbot is not programmed to provide explicit or categorical answers to your questions. Rather, it transforms your questions into API requests that are distributed across the various IT services officially administered by EPFL. Its purpose is solely to collect and recommend relevant references to content that you can explore to help you answer your questions.

Description

This lecture covers Data Denoising Diffusion Models, which consist of a forward diffusion process adding noise to input and a reverse denoising process generating data. It explains the formal definitions of the forward and reverse processes, the diffusion kernel, and the training objectives. The lecture also discusses the parameterization of the denoising model, progressive distillation for faster sampling, classifier guidance and classifier-free guidance techniques. It concludes with the challenges of applying diffusion to text and summarizes the key points of diffusion for data generation and denoising.

Official source

Related concepts (32)

Sampling (statistics)

In statistics, quality assurance, and survey methodology, sampling is the selection of a subset or a statistical sample (termed sample for short) of individuals from within a statistical population to estimate characteristics of the whole population. Statisticians attempt to collect samples that are representative of the population. Sampling has lower costs and faster data collection compared to recording data from the entire population, and thus, it can provide insights in cases where it is infeasible to measure an entire population.

Nonprobability sampling

Sampling is the use of a subset of the population to represent the whole population or to inform about (social) processes that are meaningful beyond the particular cases, individuals or sites studied. Probability sampling, or random sampling, is a sampling technique in which the probability of getting any particular sample may be calculated. In cases where external validity is not of critical importance to the study's goals or purpose, researchers might prefer to use nonprobability sampling.

Text editor

A text editor is a type of computer program that edits plain text. Such programs are sometimes known as "notepad" software (e.g. Windows Notepad). Text editors are provided with operating systems and software development packages, and can be used to change files such as s, documentation files and programming language source code. Plain text and Rich text There are important differences between plain text (created and edited by text editors) and rich text (such as that created by word processors or desktop publishing software).

Survey sampling

In statistics, survey sampling describes the process of selecting a sample of elements from a target population to conduct a survey. The term "survey" may refer to many different types or techniques of observation. In survey sampling it most often involves a questionnaire used to measure the characteristics and/or attitudes of people. Different ways of contacting members of a sample once they have been selected is the subject of survey data collection.

Text file

A text file (sometimes spelled textfile; an old alternative name is flatfile) is a kind of that is structured as a sequence of of electronic text. A text file exists stored as data within a . In operating systems such as CP/M and MS-DOS, where the operating system does not keep track of the file size in bytes, the end of a text file is denoted by placing one or more special characters, known as an (EOF) marker, as padding after the last line in a text file.

Related lectures (64)

Approximate Query Processing: BlinkDB

Introduces BlinkDB, a framework for approximate query processing using sampling techniques.

Generative Models: Self-Attention and Transformers

Covers generative models with a focus on self-attention and transformers, discussing sampling methods and empirical means.

Wireless Receivers: Time and Phase Offset

Covers the impact and compensation of time and phase offset in wireless receivers.

Generative Models: Boltzmann Machine

Covers generative models, focusing on Boltzmann machines and constrained maximization using Lagrange multipliers.

Fourier Transform and Sampling