LG Tech Conference 2023: Where LG Technology Meets the Best R&D Talents

The LG Tech Conference, where the top talents in R&D come together, took place on March 16 at LG Science Park in Magok, Gangseo-gu, Seoul, the center of LG R&D. The LG Tech Conference, which was first held in 2012, is an event where LG's technological innovation status and vision are presented, and it serves as a venue for mutual networking between talents. Those attending the event can experience the technology of the LG industrial field and interact with fellow talents who are interested in related technologies. What’s more, it’s an opportunity for executives and employees of affiliates, including top management, to communicate with talents and meet excellent R&D personnel.

This LG Tech Conference attracted even greater expectations and attention by enhancing the Tech Session to showcase the specialized technologies of LG affiliates. The conference had previously been held as a recruitment-focused event, but it recently underwent a reorientation to become a more technology-focused conference.

The LG Tech Conference 2023 was full of things to see and useful information. Let’s have a look at the highlights.

Announcing the Grand Start of the LG Tech Conference: Opening

The opening began with a video of LG Electronics’ robot CLOi. CLOi introduced the exhibition halls in which participants could directly see the technology being developed by LG.

Next, LG AI Research’s AI Artist Tilda appeared to share the experience of showcasing around 200 outfits it designed with designer Park Younhee at New York Fashion Week. “There are seniors waiting to hear your concerns that couldn’t be resolved in the Tech Session. Everything’s off the record there. Ask anything you want,” said Tilda, urging participants to fully utilize the Networking Session.

As it finished introducing the events, Tilda welcomed Chairman and CEO of LG Corp., Kwang Mo Koo to the stage. He started by introducing a drawing Tilda drew with the keywords dream, future, and people.

He went on to explain that, “Dream, future, and people are the keywords I want to talk to you about today. The dream of LG is to make the world a place where everyone can smile by bringing joyful experiences into people’s lives and transforming imagination into a better future. In order to realize this dream, we value technology and innovation, and the people and talents that make it possible, and this is a principle that has been cherished in LG’s history for more than 75 years.”

He also added that, “As you take a look at the various future innovation technologies that LG is leading, we hope that you’ll be able to feel how these technologies are being made into reality in the industrial field and how various talents come together to create new values and grow together with the company. I hope this conference helps orient you in the right direction to achieve your dreams and grow in the future.”

Experiencing LG’s Technology and Vision in One Place: Tech Session – Tech Session

During the Tech Session, LG affiliates shared their technological development achievements, tasks, and visions in various fields such as AI, Metaverse/Blockchain, Software, Smart Factory, Big Data, and more. Among them, not only LG AI Research but also LG Electronics, LG Uplus, LG Energy Solution, LG Innotek, LG CNS, and LG Display participated in eight presentations related to AI.

From LG AI Research, Seunghwan Kim, Leader of the Vision Lab, started off by presenting “LG’s Multimodal Expert AI Research.”

EXAONE, a multimodal AI developed by LG AI Research, has a two-way feature that creates an image that matches text entered by the user and creates text that explains an image in detail when an image is entered. Kim explained the technological differences of EXAONE in image-to-text generation and text-to-image generation technologies.

Regarding image-to-text generation, he explained that, “Existing image captioning models can’t process images outside the range of data used for training. However, EXAONE’s learning data, which consists of more than 350 million images, includes various types of images such as photos, illustrations, graphics, and more, and the text describing each image also contains all the information about the details of the image.”

He also emphasized that, “Understanding the text was greatly improved with the use of UniCLIP, a multimodal representation model presented at NeurIPS last year, and the EXAONE language model together as text encoders. In addition, we’re developing technologies that can create images in a style that users prefer through various inputs rather than creating images only from texts.”

Next, Sehui Han, Leader of the Materials Intelligence Lab, gave a presentation on “How could AI transform chemical research?” First, she introduced DDU technology that collects material data by reading chemical structural formulas and table data in papers, then explained that, “There are chemical structure formulas in papers or patents that are referenced often in the field of chemistry. Through a vision model, we need to convert these into something a computer can understand. Other information written in tables must also be read and shaped into a form that we can use. After nearly two years of research, we were able to develop a model at the level of achieving global State-of-the-art (SOTA) in a language-based content extraction competition in the chemical biomedical field.”

The material data collected with this method is used to train the molecular design models. The structure determines the physical properties, and the goal is to extract this decision logic into an AI model.

Han emphasized that, “If an AI model collects a lot of structure and property data and learns from it, it’ll end up having logic. With this logic, since physical properties come from the structure, finding new structures by entering the desired physical properties is only a matter of time. With this technology, processes that once took around 3.5 years for new drug projects can be shortened to 8 months.”

In other AI Tech Sessions, presentations that were given were as follows: LG Electronics’ Hoyoung Choi, AI Platform TP Leader of the Ambient Intelligence Lab, “ATOM - Ambient Computing Technology of Multiverse”; LG Uplus’ Sanghyeon Lee, Senior of the Recommender Systems & Personalization Team, “AI Recommendation/Search Technology for Maximizing the Customer Media Experience”; LG Energy Solution’s Junho Yim, AI Technology Team Leader, “AI Technology and Application Cases to Minimize the Outflow of Defects in Smart Factories”; and LG Innotek’s Seokbok Jang, Research Fellow, AI/BigData Solution Team, “Introduction of AI Technology for the Intelligent Manufacturing of Precision Parts.”

Meanwhile, from LG CNS, General Consultant Sanggyo Jeong of the AI Engineering LAB introduced “Bringing AI from the Lab to Operation, ‘SW with AI,’” while Byungseung Lee, Managing Director at LG Display DX Group, introduced “The Future of Intelligent Autonomous Factories Dreamed of by LG Display and Challenges That Come With It.”

A Place for In-depth Exchange Between Experts: AI Poster Session and Networking Session

Once the AI announcement within the Tech Session was over, the AI Poster Session began. Authors of papers, including papers recently accepted at international academic conferences, were able to introduce their research to visitors by setting up posters on AI research on various topics. At the AI Poster Session, LG AI Research, LG Uplus, and LG Electronics participated and presented papers.

[ISMIR 2022] Symbolic Music Loop Generation with Neural Discrete Representations: Data Intelligence Lab, Sangjun Han (Link)

In this paper, research on generating music that sounds like it was composed by humans using AI was introduced. Sangjun Han of the Data Intelligence Lab explained that, “In the case of language models, sentences are made longer, but in the case of music, as patterns are repeated, making it long doesn’t make much sense. So, we thought it would be more meaningful to make short 8-measure loops. It’s easy for humans to know which part of a piece of music is a loop, but this is difficult for machines. Loops were extracted using deep learning, and music was generated to evaluate how similar they were to music composed by humans.”

[NeurIPS 2022] Transferring Pre-trained Multimodal Representation with Cross-modal Similarity Matching: Advanced ML Lab, Byungjip Kim (Link)

To achieve zero-shot performance by properly recognizing unseen data in the large multimodal model, it must be trained with a large amount of data or resources. This paper proposes a method for achieving zero-shot performance by re-using logic from already well-learned and high-performing models. Byungjip Kim of the Advanced ML Lab explained that, “We made the method follow exactly how a well-learned model, in other words, the teacher, thinks when images are entered, as well as how the model thinks of the relationship between texts and images by entering texts together with images. As a result, we needed 400 million images with conventional methods to achieve the same performance, but the method presented in this paper used only 12 million sheets and 3% of the data to achieve the same performance.”

In the Networking Session that adorned the end of the event, researchers from each affiliate, who served as Tech Session presenters, attended and had time to talk freely in a comfortable atmosphere, answering questions from participants and providing career counseling.

At the LG Tech Conference 2023, held in the spring, excellent talents in science and engineering gathered to network and got to know LG's top-notch staff and technology. LG AI Research introduced its most notable technologies through the Tech Session and Poster Session at this LG Tech Conference, demonstrated its research capabilities, and communicated with top talents. LG AI Research plans to continue meeting AI talents who will take the lead in Korea while producing excellent research results.

▶ LG Tech Conference 2023 Video Clip (Link)