LG AI researchers are entering the domain-specific AI platform service market by unveiling the massive AI 'ExaOne 2.0'.

On the 19th, LG AI researchers held the 'LG AI Talk Concert 2023' at LG Science Park in Magok, Gangseo-gu, Seoul, and unveiled the ultra-large multimodal AI 'EXAONE 2.0'. Universe, Discovery, and Atelier, the three major platforms of EXAONE that serve as the foundation for developing 'expert AI' services, were introduced.

Since its initial unveiling in December 2021, 'ExaOne' has been trained on 45 million professional documents and 350 million images, with the amount of patents and training data increasing by more than four times compared to existing models. It was developed as a bilingual model capable of understanding and responding in both Korean and English simultaneously.

Through lightweighting and optimization, the inference processing time was reduced by 25% and memory usage by 70% while maintaining the same performance as the existing model. The multimodal model doubled memory usage but reduced inference processing time by 83%.

The main evolution of 'ExaOne 2.0' is that the model size, type (language, vision, multimodal), and language used can be customized to suit the desired purpose or budget.

ExaOne will be applied as a chatbot in LG Electronics' AICC (AI Contact Center). AICC is scheduled to transition to a formal service in the second half of the year and will be expanded to English-speaking countries starting next year.

ExaOne Universe has integrated functions such as Q&A and conversation, text classification and summarization, keyword extraction and generation, and translation into an expert chatbot AI platform. ExaOne Universe generates responses by including not only pre-trained data but also the latest specialized data for each domain. Currently, specialized services for each of ExaOne Universe's professional domains, including chemistry, biotechnology, pharmaceuticals, healthcare, finance, and patents, are also being prepared.

ExaOne Discovery is a chemical and bio platform applied to the exploration of new materials, new substances, and new drugs. It is expected that development time can be reduced from the existing 40 months to 5 months using this platform, which utilizes 'Deep Document Understanding (DDU)' technology.

LG AI Research Institute showcased a process of linking ExaOne Universe and ExaOne Discovery to find candidate materials and predict synthesis results by querying AI, under the theme of developing additive materials that can be used in the development of eco-friendly batteries.

ExaOne Atelier is a multimodal AI platform specialized in image generation and image understanding, based on ExaOne 2.0, which was trained on 350 million copyrighted image-text pairs.

LG AI Research is continuing joint research and development with Parsons School of Design in the U.S., and has also incorporated the 'Captioning AI' feature, which was commercialized with Shutterstock last June, into ExaOne Atelier.

At the event, Bae Kyung-hoon, Head of LG AI Research Institute, emphasized, “We will develop into a globally competitive AI company that creates ‘customer value differentiated from other generative AIs’ by building tangible success stories through cooperation with domestic and international partners.”