LG AI Research 345

LG AI Research attended the world-renowned AI conference ICML 2023, presenting various research achievements and sharing the latest technology trends.

One of the key research topics, Unlearning, is a technique for efficiently removing information about specific data from pretrained models such as GPT-3 or Stable Diffusion. The Data Intelligence Lab proposed a Gradient Surgery technique that can remove specific instances at minimal cost while maintaining the performance of generative models.

Instead of the existing complex Hessian computation method, this technique effectively removes data by updating model parameters in a direction that minimizes the impact on generated outputs.

Research in the field of Causal Discovery, which finds causal relationships from data, was also introduced. Recently, research that goes beyond existing statistical learning to efficiently find complex causal structures using Bayesian Inference or GFlowNet has been drawing attention.

In addition, LG AI Research is focusing on applied research in next-generation core AI fields such as Causal Inference, Geometric Deep Learning, and Large Scale Foundation Model.