LG AI Research Develops Foundation Model Using Amazon SageMaker

South Korean conglomerate LG Group collects vast amounts of data from its companies, which include home appliances, telecommunications, batteries, and pharmaceuticals. A key pillar of the group’s digital transformation is developing AI technology and integrating AI into its products and services. The group established LG AI Research to harness the power of AI in its digital transformation strategy, develop better customer experiences, and solve common industry challenges.

When LG AI Research decided to develop its next-generation foundation model, which takes inspiration from how the human brain works and has an advanced capacity for learning and making judgments, it searched for the most efficient machine learning (ML) platform to handle vast amounts of data and large-scale training and inference. The foundation model needed to train on dozens of terabytes of data to make human-like deductions and comprehend texts and images. Moreover, the project required a high-performance compute infrastructure and the flexibility to increase the number of parameters to billions during training.

Workflow automation was also important, as multiple models or downstream tasks needed to be completed simultaneously. To meet these requirements, the institute looked at an on-premises infrastructure, but costs were too high, and it would require 20 employees to configure and maintain the on-premises hardware. It would also require upgrading the GPUs every year and adding more GPUs to support workload spikes. Considering all the challenges in an on-premises solution, LG AI Research decided that Amazon SageMaker was the best fit for this project.

LG AI Research successfully deployed its foundation model, EXAONE, to production in one year. EXAONE, which stands for “expert AI for everyone,” is a 300-billion-parameter multi-modal model that uses both images and text data.

LG AI Research used Amazon SageMaker to train its large-scale foundation model and Amazon FSx for Lustre to distribute data into instances to accelerate model training. By building on AWS, LG AI Research was able to resolve issues, implement checkpoints, fine-tune, and successfully deploy the model to production.

LG AI Research’s Gwang-mo Song explains, “By using Amazon SageMaker’s high performance distributed training infrastructure, researchers can focus solely on model training instead of managing infrastructure. In addition, by leveraging the parallel data library from Amazon SageMaker, we could obtain training results quickly as the number of GPUs and model parameters increased.”

LG AI Research reduced costs by approximately 35 percent by eliminating the need for a separate infrastructure management team. It also increased the data processing speed by about 60 percent using the Amazon SageMaker distributed data parallel library.

Using EXAONE, LG AI Research developed an AI virtual artist called Tilda. The fundamental power of Tilda’s artistic qualities comes from EXAONE, which was trained using 600 billion pieces of artwork and 250 million high-resolution images accompanied with text. The virtual artist created 3,000 images and patterns for fashion designer Yoon-hee Park, who designed more than 200 outfits for the 2022 New York Fashion Week using Tilda’s images and patterns.

Park’s work with LG AI Research demonstrated the potential of expanding AI technology to the art industry, growing the AI ecosystem and fostering cross-industry collaboration. The company recently announced a partnership with Parsons School of Design in New York City to conduct joint research on advanced AI technologies to leverage in the fashion industry.

With Tilda, EXAONE has shown how foundation models can be used to transform a wide range of sectors, from manufacturing and research to education and finance. LG AI Research continues its work to make human life more valuable using its foundation model and looks forward to collaborating closely with AWS on future projects.