AI BriefingLG AI Research 191
Key point
This review focuses on the Diffusion Model, one of the most-noticed generative models among the diverse works presented at NeurIPS 2021, and covers the key research on it.
Details
Researchers at LG AI Research's Applied AI Research Lab have summarized major research papers presented at NeurIPS 2021. The core of this review is the Diffusion Model, one of the most actively researched generative models in recent times.
Generative models estimate the distribution of data to generate new data, and are broadly divided into Discriminative and Generative models. This article compares and analyzes the characteristics of various generative models.
- Auto-regressive models (ARMs): Use the conditional probability of sequential variables and can compute the likelihood of data, but sampling speed is slow.
- Variational Autoencoders (VAEs): Utilize latent variables to enable fast training and sampling, but have complex computations and a restrictive prior distribution.
- Generative Adversarial Networks (GANs): Generate high-quality samples through competition between a generator and a discriminator, but training can be unstable.
- Diffusion Model: Learns the process of gradually adding noise to data and then recovering it. It estimates complex distributions well, and is drawing attention recently for generating higher-quality samples than GAN-based models and showing better performance than ARM.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.