LG AI Research and AIIS Team Up at the 2022 AIIS Spring Retreat
Leading Korean AI organizations meet at the SNU AIIS Retreat
On April 15, the 4th AIIS Spring Retreat was held at the Seoul National University (SNU) Engineer House. At this year’s AIIS Spring Retreat, the distinguished researchers of SNU and various industry-leading AI companies got together to share the latest AI research while actively interacting with each other. The AIIS Retreat, meaning to withdraw and take a rest during the war, is a special event where AI companies that usually are in fierce competition gather to discuss core AI issues and trends. In addition to major members of the Artificial Intelligence Institute Seoul National University (AIIS) such as LG AI Research, NAVER, and Kakao Enterprise, Megazone Cloud, GC Pharma, and CJ AI Center newly joined the event this year, adding up to a total of six member companies.
Scene of the 4th AIIS Spring Retreat
This year's AIIS Retreat was divided into the “oral session” where researches conducted by each organization were presented and the “poster session” where researchers made posters using their research outcomes to communicate freely with each other.
Prior to the start of the main event, AIIS Director Byoungtak Zhang expressed his anticipation for this event by stating, “AI is not something exclusive to the science and engineering field. It requires pan-business cooperation encompassing various industries. We hope this year's event serves as a great opportunity to exchange many information on AI with a broader perspective and create the platform for synergy between AI companies and academia.”
Gyeongmu Lee, Dean of the Graduate School of AI at SNU who hosted this event, emphasized the meaning of this event in his congratulatory speech saying, “It is crucial to construct an organic cooperative system among relevant organizations such as educational institutes, government, and businesses. From this perspective, the AIIS Retreat is a meaningful event as leading Korean AI companies and individual researchers can meet face-to-face and freely communicate about AI.” He also expressed his gratitude to the member companies including LG AI Research by commenting, “The growing size of the retreat with the passage of time and the fact that concrete research outcomes are coming out was made possible because of the member companies.” Dean Lee also stressed the outcome of a total of 95 papers being published, including 70 top-tier academic papers since last year, and stated his aspiration to increase this by 1.5 times in 2022.
The congratulatory speech was followed by corporate presentations of a total of six member companies including LG AI Research.
Woohyung Lim, Applied AI Research Lab Leader, presenting at the 4th AIIS Spring Retreat
Applied AI Research Lab Leader Woohyung Lim introduced the role of LG AI Research within the LG Group as well as researches currently being conducted and research achievements in the first year of its founding. Lim also shared future plans for AI by the company and examples of applying the technology in the field. In particular, he expressed a strong will to interact with relevant companies and academic circles, “In order for the goal of LG AI Research to be achieved, it is important to reflect your research in businesses. For this, LG AI Research is expanding its cooperation with various alliance partners and plans to actively pursue collaboration with academic circles such as AIIS.”
In addition, LG AI Research and AIIS agreed to establish a joint AI research center in February to conduct closer joint research, undergo super-giant AI model research, and hold AI global forums in unison. “While quality research infrastructure is important to conduct good research, I believe that the most important element is having good colleagues. In addition to operating various fellowship programs and establishing overseas research labs, LG AI Research will continue its efforts to construct an environment for pursuing joint research with good people.” The presentation was ended with the introduction of the strengths of LG AI Research.
LG AI Research making great strides with diverse research
LG AI Research has been achieving outstanding research outcomes with the challenging goal of creating AI that can be applied in everyday life to create a better life for humanity. In the first year of its establishment, 18 papers were accepted by global top-tier academic journals in addition to a variety of other papers that were presented. At this event, three papers were introduced at the oral session and poster session.
Oral session
[CVPR 2022] L-verse: Bidirectional Generation between Image and Text — Vision Lab, Taehoon Kim (Link)
Taehoon Kim, presenting at the 4th AIIS Spring Retreat oral session
The bidirectional image text generation algorithm L-Verse was introduced in the oral session. It allows simultaneous bidirectional generation work unlike the existing single-direction generation such as text-to-image or image-to-text. Furthermore, images were grouped by patch to treat as a single word in order to minimize information loss that occurs when an image or text passes through an encoder. By comparing this with the most similar word per patch, it could generate detailed captions or images, thereby maximizing its performance.
Poster session
[CVPR 2022] Instance-wise Occlusion and Depth Orders in Natural Scenes — AI Human Company, Hyunmin Lee (Link)
Hyunmin Lee, presenting at the 4th AIIS Spring Retreat poster session
The Insta Order model that was introduced in the first poster session is a study that began with the question on whether existing models could overcome the limitations of predicting the image by pixels through predictions based on identifying the relationship between objects. By comparing InstaOrder Net and InstaDepth Net with existing datasets, it was evident that compared to the existing SOTA model, it had better performance.
[AAAI 2022] Neural Marionette: Unsupervised Learning of Motion Skeleton and Latent Dynamics from Volumetric Video — Vision Lab, Jinseok Bae (Link)
Jinseok Bae, presenting at the 4th AIIS Spring Retreat poster session
Lastly, research on neural marionette was presented at the poster session. Neural marionette is a model that excludes all prior knowledge of a structure and only uses raw data of a skeleton captured by camera to learn the structure of motion dynamics. This model is capable of conducting three tasks all at once: (a) motion generation that creates motion based on information on learned motions; (b) motion interpolation that reinforces motion between two separated poses; and (c) motion retargeting that transfers the motion to a stopped 3D target data by extracting the motion using source observation.
In addition to the contents presented by LG AI Research, outstanding research achievements of various labs were introduced at the 2022 4th AIIS Spring Retreat. As AI technologies are being developed and applied in various areas, it was an opportunity to feel the passion of the participating companies toward AI research as they overcome limitations in AI by resolving various problems with their technologies. LG AI Research plans to continue to actively cooperate through interaction with academic and industrial circles such as the AIIS Spring Retreat.