AI BriefingLG AI Research 302
Key point
LG AI Research reviewed major research trends in generative models, LLMs, and reinforcement learning presented at NeurIPS 2022.
Details
At the world-renowned AI conference NeurIPS 2022, LG AI Research presented new papers and technical demos while identifying global research trends. This review covers notable papers in the fields of Generative Modeling, large language models (LLM), and Reinforcement Learning.
The Score-based Generative Model (SGM), a core method in recent generative modeling, has established itself as the standard for image synthesis and text-to-image modeling. However, its slow generation speed due to the multi-step sampling process has been pointed out as a limitation, and research to improve this is actively underway.
Key research examples are as follows:
- DPM-solver: Defines the reverse process of SGM as a probability flow ordinary differential equation (ODE) and increases sampling speed by approximating only the nonlinear term.
- GENIE: Proposes a higher-order solver based on Taylor's method, and significantly reduces the computation of the inference step through a Distillation approach that adds a small neural network on top of the existing model.
- INDM & Riemannian SDE: Research that improves image generation performance by defining diffusion in Latent space or utilizing Riemannian geometry also drew attention.
In addition, it was shown that SGM can be applied in a Plug-and-play manner to visual downstream tasks such as Super-resolution or Inpainting without any additional training.
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.
Related
Explorative Modeling: Learning from the Best of K Guesses