AI BriefingLG AI Research 349
Key point
LG AI Research unveiled an efficient unlearning technique to protect data privacy in generative models at ICML 2023.
Details
LG AI Research participated in ICML 2023, a world-class AI conference, and presented research results aimed at addressing data privacy issues in generative AI models.
In particular, the company focused on Unlearning technology. Unlearning is a technique that efficiently removes specific personal information or copyright-related data from data that models such as GPT-3 or Stable Diffusion have already learned. The key is to eliminate the influence of specific data at minimal cost while maintaining the model's performance as much as possible.
In this research, LG AI Research proposed a method that minimizes the impact on the output of generative models by utilizing the Gradient Surgery technique. In complex DNN (Deep Neural Network) and generative model environments where the existing Newton Update method is difficult to apply, the company proved both theoretically and experimentally that the influence of data can be effectively removed by minimizing gradient conflicts.
LG AI Research plans to further refine the boundaries of privacy protection in the future by introducing a weight-search framework based on Differential Privacy.
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
Google Introduces Private, Persistent Memory to Private AI Compute PlatformComfy Launches Comfy Router: A Unified API for Frontier Media Models