LG AI Talk Concert 2023, Large-Scale AI EXAONE 2.0 Unveiled for the First Time

LG’s large-scale AI model EXAONE 2.0 was unveiled for the first time at the LG AI Talk Concert 2023 held on Wednesday, July 19, 2023. Attendees were introduced to the advanced performance and feature of EXAONE 2.0 as well as the commercial use it offers. We invite you inside.

EXAONE 2.0, an advanced large-scale multimodal AI

LG AI Talk Concert 2023 began with opening remarks from Kyunghoon Bae, Chief of LG AI Research. When speaking about the development of EXAONE, Chief Kyunghoon Bae said, “Ever since LG AI Research was founded in December 2020 with the dream of creating a better life through AI, we have focused on solving the challenges that exist across various industries. Using our progress within these industries as a foundation, we developed something even more innovative: EXAONE, an expert AI meant for everyone.”

From its inception, our goal with EXAONE has been to revolutionize the industry through the comprehension of multimodal data, which includes both linguistic and visual data. Furthermore, we are taking on the challenge of making new scientific discoveries based on the cumulative knowledge of mankind. LG AI Research is the only company in Korea that is developing generative AI that can understand and produce language and images and comprehend expert-level documents, and it is also among just a handful of companies in the world offering this kind of AI.

However, there are a few issues that currently make the application of generative AI challenging in actual industrial settings. The level of confidence in the results is not high; the process of training and operating a large-scale model is very expensive; and there are concerns about potential leaks of internal data.

To address these issues, LG AI Research focused on improving the quality of training data, increasing cost efficiency, and providing customized models. This led to a game-changing development in the world of generative AI: the new EXAONE 2.0.

1) EXAONE 2.0 – Data aspects

EXAONE 2.0 has secured high-quality data centered around specialized data. It holds more than 45 million papers and pieces of patent data and more than 350 million pieces of image-text data, and we resolved all copyright and privacy issues before proceeding with training. Furthermore, in order to overcome the limitations of a Korean dataset, we created a bilingual model that is capable of understanding and generating responses in both Korean and English. By increasing the amount of Korean-English training data to more than four times than before, we were able to dramatically improve performance through synergy between Korean and English data. In addition, by using bilingual multi-task tuning, we were able to maximize the performance of EXAONE.

2) EXAONE 2.0 – Increasing cost efficiency

By implementing EXAONE 2.0’s large-scale language model, we were able to reduce the time required to make an inference by 25% and decrease memory usage by 70% when compared to our previous model. By doing so, we were able to achieve the same level of performance as our previous model with a 78% reduction in costs. In addition, by doubling the memory usage of our multimodal model to increase the quality of its results, we were able to see an 83% decrease in inference time, leading to a 66% reduction in costs. We improved the economic aspect that had been pointed out as one of the limitations of large-scale AI.

3) 3) EXAONE 2.0 – Provide custom models tailored to customer needs

Last but not least, EXAONE 2.0 will be available as a customized model to suit our clients’ needs and use cases. Clients will be able to choose text or images as their inputs and outputs and pick the language. We can also fine-tune our clients’ internal data according to their field of domain expertise and allow them to select configurations. Since this custom approach can maintain domain-specific expertise while resolving internal data security concerns, we expect that the use of EXAONE in specialized industries will increase greatly.

The 3 Platforms of EXAONE 2.0 – EXAONE Universe, EXAONE Discovery, EXAONE Atelier

Next came the introduction of the three major platforms featuring EXAONE 2.0: EXAONE Universe, EXAONE Discovery, and EXAONE Atelier. EXAONE Universe offers reliable inferences based on expert knowledge, while EXAONE Discovery specializes in exploring new scientific facts. EXAONE Atelier is a service platform that provides creative inspiration and ideas. Let’s take a closer look at each platform.

Providing expert-level answers and insights with EXAONE Universe

EXAONE Universe is able to immediately generate expert-level answers to questions through a large-scale language model, which is the core of generative AI technology. It is also a platform that researches the latest technology and helps companies with their professional tasks. We focused on creating a model with comprehensive reasoning and logical thinking based on expert documents and the most up-to-date knowledge that could generate fluent answers through the collection of factual evidence.

Moontae Lee, leader of LG AI Research’s Advanced ML Lab, gave a detailed introduction to EXAONE Universe. When explaining the purpose behind the development of EXAONE Universe, Moontae Lee said, “Even now, in 2023, most professionals use search engines to look up information. However, the process of finding the most relevant documents from the search results, extracting the necessary information from each one, and combining everything into something useful takes a considerable amount of time.”

Current language models generate the next word that is appropriate based on context well, but they struggle with forming an overall picture of the basis for the responses, and there are frequent instances of hallucinations, which occur when statements that are not true are made.

To solve such issues, we have designed EXAONE Universe with a comprehensive and logical reasoning function that identifies expert documents from the latest database, makes detailed inferences based on what is found within these documents, and then synthesizes these different pieces of information into insights. As such, it can naturally refer to the reasoning used to form its response.

Moontae Lee demonstrated the EXAONE Universe on-site and explained its features in detail. First, we asked it the question “How do you overcome hallucination in a large language model?” and looked at the responses given. EXAONE Universe offered a method of using both text and visual data and NLI(Natural Language Inference) techniques and gave in-depth responses based on each factor.

Next, we asked the questions “What should we study generative AI?” and “Will AI and automation replace human workers?” “What should we study generative AI?” was a deliberately written question, but EXAONE Universe delivered a surprising result by comprehensively inferring “why” and “what” to generate its response.

Exploring new scientific facts with EXAONE Discovery

Next came the introduction to EXAONE Discovery. EXAONE Discovery is a service platform that explores new scientific facts, and the first task it is taking on is the development of new materials and drugs with AI. In the first stage, the understanding phase, we developed and commercialized technology that understands detailed documents, extracting data and knowledge from expert literature such as papers and patents. Then, through EXAONE 2.0, we are advancing to the design-prediction stage, the second stage, which involves designing a molecular structure with desired properties and predicting the result of intermolecular synthesis.

Sehui Han, leader of the Materials Intelligence Lab, explained EXAONE Discovery in detail. At the beginning of the explanation, Sehui Han said, “EXAONE Discovery is a platform that can serve as a GPS in the process of discovering new materials and drugs.”

Sehui Han also introduced EXAONE Discovery’s vision and stated, “Up until now, designing and synthesizing molecules was a task reserved for human scientists. But going forward, AI will be able to help scientists quickly design the substances that they’re looking for, and AI models will be able to easily predict the behavior of a substance without physically synthesizing and testing it in a lab. In other words, we’re looking at the beginning of a new era in which AI will be able to rapidly develop new materials with only a small amount of data.”

Using existing research methods, researchers would analyze past results, form a hypothesis based on this data, and then physically create the material and examine it in order to test their hypothesis. During this process, countless experiments are conducted, failures occur, and large amounts of time and money are consumed.

EXAONE Discovery’s services will drastically streamline this process. By inputting just one sentence, the model will go over various references and complex details to suggest the optimal means of developing a desired material. This will result in scientists minimizing the risk of failure in their research.

EXAONE Discovery also showed off its capabilities through a demonstration. Influenced by the recent environmental issue of reducing fluorine content in battery additives, we searched for new materials to replace existing battery additives using EXAONE Discovery.

Providing creative inspiration as well as original ideas with EXAONE Atelier

Finally, EXAONE Atelier, which aims to provide creative inspiration as well as original ideas, was introduced. EXAONE Atelier has been trained on a data set of 350 million image-text pairs and excludes any data that could potentially infringe on copyright.

By using EXAONE Atelier, designers will be able to create new images that previously did not exist, enabling them to bring their design concepts to life. With EXAONE Atelier, you will be able to discover fun new ideas that you might not have thought of and use them in your design work, minimizing the amount of time it takes to create a product.

EXAONE Atelier was introduced by Seunghwan Kim, leader of the Vision Lab. Seunghwan Kim explained EXAONE Atelier’s zero-shot image captioning technology and emphasized the fact that the AI goes beyond just describing inputted images in a natural language like a human; it is able to perform tasks desired by the user by combining with the EXAONE language model.

To see EXAONE Atelier in action, we showed LG product photos and asked it to create marketing phrases. EXAONE Atelier generated these phrases: “Introducing the ultimate cosmetic set designed to take care of your skin and enhance your beauty. Enclosed in a specially formulated blue box package, this set is tailored to improve your appearance, leaving you with a radiant glow that boosts your confidence.” This creative marketing copy showcased a comprehension of the features shown in the images of the cosmetic package.

What’s more, in addition to the marketing copy, we showed examples of post captions being generated when users upload their everyday photos onto their social media feeds and a video showcasing how an inventive story that appeals to children can be made using pictures.

Q&A session with various insights

After the introduction of EXAONE 2.0, there was a Q&A session. In response to pre-collected questions from media outlets as well as on-the-spot questions, Kyunghoon Bae, Chief of LG AI Research, along with Seunghwan Kim, leader of the Vision Lab, Jungkyu Choi, group leader of Multimodal AI Group, Hwayoung(Edward) Lee, head of the AI Business Development Unit, Sehui Han, leader of the Materials Intelligence Lab, and Moontae Lee, leader of the Advanced ML Lab, gave in-depth answers based on their respective specialties. There were so many questions that it wasn’t possible to answer all of them during the time allotted.

There was also an on-site booth where attendees could try out the EXAONE 2.0 platform. Attendees were able to experience the EXAONE Atelier and its amazing technological capabilities for themselves.

LG AI Research’s EXAONE is pioneering and achieving success across various industries through partnerships with domestic and global partners as well as LG affiliates. Going forward, LG AI Research will lead in popularizing generative AI technology and become an AI research and development company that is competitive not just in Korea but also in the global market.

If you want to know more about the presentations at the LG AI Talk Concert 2023, check them out at the YouTube link below.

▶ Watch the full version of the LG AI Talk Concert 2023 (Link)