In commemoration of its second anniversary, LG AI Research hosted the LG AI Talk Concert online under the topic of “Challenges and Innovation for Expert AI Development” and reflected on its accomplishments in 2022. The event was centered on super-giant AI EXAONE, introduced by LG AI Research in December 2021.

Image 1. Kyunghoon Bae, Chief of LG AI Research, introducing LG's super-giant AI EXAONE

LG's Expert AI aims to identify areas where AI excels, acts as a top-one-percent expert in the field, and collaborates with humans to create unparalleled customer value. Based on these objectives, EXAONE has undergone constant evolution since its release.

EXAONE trains from academic paper and lays the groundwork for solving complex problems

LG AI Research introduced a technology in which LG's super-giant AI EXAONE trains not only the text of academic paper such as theses and patents, but also formulas, tables, and images to create a database. The platform applied with EXAONE — EXAONE Discovery — was presented during the Talk Concert. EXAONE Discovery is a platform designed based on the idea that if super-giant AI can train and apply the knowledge accumulated by humanity thus far, it can aid in the resolution of the world's most challenging problems. Commonly, AI requires human assistance to learn the jargon, formulas, tables, and images contained in scientific papers and patents. To overcome the limitations of AI, which can assist humans but cannot independently train specialized fields, LG AI Research has developed a technology that enables EXAONE to directly train not only texts but also formulas, tables, and images from academic literature such as papers and patents and convert them into a database.

In particular, Optical Chemical Structure Recognition (OCSR) technology, one of the core technologies of EXAONE Discovery, was published as a thesis. It was recently accepted at NeurIPS 2022, a global AI conference held in New Orleans, U.S. OCSR is a technology that surpasses the limitations of image captioning, which recognizes the entire molecule as a single string. It can create a database by recognizing each atom and bond type. In terms of recognition rate and efficiency, it outperforms existing AI models that can read chemical structures by more than a factor of one hundred.

AI technology utilized to address industrial and human challenges

In collaboration with LG affiliates as well as domestic and international partners, LG AI Research is also applying AI technology to actual industrial sites.

For example, LG Electronics is incorporating AI into a technology that forecasts weekly product sales demand by country and region. LG Innotek uses AI technology to increase efficiency and productivity at industrial sites, such as by recommending process conditions for centering the camera lens and sensor to reduce the process optimization time by more than 50%. It differs from other AI models in that it can gain experience while solving a variety of technical challenges in real-world industrial environments.

In addition, LG AI Research collaborates with affiliates of the LG Group to conduct diverse research on industrial applications.

The first is an investigation into a personalized anti-cancer vaccine candidate prediction model. The personalized anti-cancer vaccine is an immunotherapy that uses the patient's genetic information and cancer cell mutation information to identify neoantigens that induce cancer cell death and administer these to the patient. It was difficult to predict neoantigens due to the complex interactions between proteins and the high number of cases, despite the fact that it has been intensively studied in the bio industry and academia due to its low rate of adverse effects and recurrence. Consequently, LG AI Research has developed an AI model to predict neoantigens, which is the first step in the development of anti-cancer vaccines. In particular, compared to other prediction models used in the industry, it is anticipated to demonstrate superior performance and significantly reduce the time required for vaccine development.

In addition, LG AI Research is conducting research to develop electrolytes for lithium-sulfur (Li-S) batteries. Li-S batteries, one of the next-generation batteries, cannot be commercialized without an optimized electrolyte compound. Existing methods, such as direct experiments or simulation-based calculations, are time- and cost-intensive as the number of cases that need to be researched to find the optimized electrolyte compound is virtually infinite. LG AI Research utilizes AI at this stage to identify new solvent candidates with the highest applicability among the publicly available compounds, and is currently conducting experimental validation of these candidates.

LG AI Research is also utilizing AI to develop high-efficiency OLED light-emitting materials. Even after more than a decade of research, it is known that it is difficult to produce results in the high-efficiency light-emitting materials field as it requires finding the material with the best performance among trillions of molecules. LG AI Research has successfully developed an AI model that predicts luminescence performance, and is currently conducting synthesis evaluation using the identified promising candidate molecules.

Super-giant language model that offers more convenient services

Image 2. EXAONE language model introduced by CSAI Honglak Lee

LG AI Research also disclosed the outcomes of its research in the new lightweight/optimized AI technology applied to a super-giant language model. It actively reflected customers' voices that it is burdensome to use super-giant AI in the industrial field as it requires a great deal of time and resources for additional training and investment for AI model development.

As a result of customer-focused research conducted over the past year, LG AI Research has applied lightweight technology as well as optimization technology for each field. The super-giant language model applied with the new technology introduced at the Talk Concert decreased GPU usage by 63% compared to EXAONE unveiled last year. It also increased inference speed, which determines the speed of AI development, by 40% and achieved SOTA (State-of-the-Art) level accuracy.

In addition, the size of the super-giant AI model has been configured in a variety of ways. Customers can use it more conveniently according to the level of difficulty and intended use, such as for simple, moderately difficult, and extremely difficult problems.

LG AI Research's super-giant language model is actively used in real industrial settings, including AI Banker developed by Woori Bank for customer consultation and banking business; LG Household & Health Care's AI Contact Center (AICC), an intelligent customer response service; and app store customer review analysis developed by LG Uplus to quickly respond to customer reviews about applications. Other services are being developed for a variety of other industries. LG AI Research intends to support LG affiliates and global partners by incorporating new technology into EXAONE, which can develop AI expert services for customers by training additional specialized languages in each field.

New EXAONE platform utilized in various fields

Image 3. EXAONE platform “EXAONE Atelier” introduced by Vision Lab Leader Seunghwan Kim

LG AI Research has also introduced a platform that enables users to employ and utilize super-giant AI without the need for coding expertise. This is significant in that it has contributed to the expansion of the ecosystem by creating a service platform that allows non-AI experts to utilize super-giant AI.

The “EXAONE Universe” is a super-giant language model-based platform designed for businesses that wish to provide services to their clients. EXAONE Universe's primary capabilities, which facilitate the development of expert AI that provides services to customers, including Q&A, dialog, text categorization/generation, keyword extraction/generation, and translation/conversion. Users can plan various services, such as AI Customer Center and AI Banker, using EXAONE Universe, from expert consultation models for each field to review analysis and creation of marketing phrases.

EXAONE Atelier is a platform that maximizes the use of EXAONE's distinctive multi-modal characteristics. Enabling bi-directional generation between text and images, it is a platform for collaborating with others to generate creative designs that have never existed in the world. LG AI Research is conducting joint research on generative AI technology using EXAONE Atelier with Parsons School of Design and Shutterstock, with future plans to create more designs by expanding the targets and scope of collaboration.

After announcing the LG AI Ethics Principles in August, LG AI Research introduced the current status of research in the fields of Fairness, Safety, and Responsible AI, in addition to the achievements obtained from previous research such as AI technology for coding and AI for composing, which convert human language into computer language.

Image 4. Universal AI developed by LG AI Research

As such, LG AI Research continues its efforts to solve tasks that required a lot of time and costs through AI. In addition, we are conducting research to implement “Universal AI,” an expert AI model that solves complex problems in the real world and facilitates optimal decision-making by utilizing global knowledge in real time. In the future, we anticipate contributing to the resolution of human problems such as diseases, energy, and resource issues, based on the knowledge and technology gained while solving industrial problems. To this end, LG AI Research intends to continue pursuing AI research that contributes to a better world by identifying superior AI talent through active investment at the LG Group level, actively collaborating with academia and industry, and establishing a more advanced research environment.

▶Watch Replay of the LG AI Talk Concert 2022 (Link)