At first, it may not be easy to understand what it means to share an AI model. How big is the AI model in real-world applications? The largest AI model currently released is a language model called PaLM (Pathways Language Model) announced by Google in April 2022, with 540 billion parameters.
In the case of the EXAM project introduced in the previous article (AI-Sharing without Sharing Data (2)), it is based on an image recognition deep learning model called ResNet-34 with 63.5 million parameters. The 20 hospitals will each learn an AI model with at least 63.5 million parameters and have a weighted average model as a global shared model. There is a magic in that the AI model is large enough to be unimaginable by our intuition. It is the power and magic of AI-Sharing, that the combined 20 large models show greater performance than each large model.
The AI-Sharing algorithms will also continue to improve in the future. Harex InfoTech's User-Centric Artificial Intelligence Research Lab newly developed an AI-Sharing algorithm IPA (Iterative Parallel Average) and ISA (Iterative Serial Average) and applied for patents. The IPA methodology performed better than the existing federated learning algorithms, and it shows that it works well even in a commercial environment where Korean language and English language data are mixed.
The results of this study will be presented at “The 23rd International Conference on Electronic Commerce (ICEC 2022)”, which will be held from June 22 to 23 at the Hotel Interburgo Daegu. According to the study, if an artificial intelligence is made from the intelligence of a person who has worked a total of 20 years, five years each at department stores in Korea, Japan, the United States, and China, existing federated learning algorithm can be compared to combining the intelligence of four people who have worked in each department store for five years. On the other hand, the IPA algorithm can be compared to the case of combining the intelligence of four people who worked at department stores in Korea, Japan, the United States, and China for one year each, then repeating this process five times to create a global intelligence. In addition, the institute is recently testing the performance of the ISA algorithm, which creates a global intelligence similar to the intelligence of a person working continuously for 20 years for department stores in Korea, Japan, the United States and China, which may lead to greater results.
Kyung Hee University's AI & BM Lab and Harex InfoTech's User-Centric Artificial Intelligence Research Lab have also found that glocalization strategy is better for AI-Sharing. It was found that rather than sharing each economic entity's whole AI model, it is better to leave the parameters of input and output parts of the AI model reflecting the specificity of each economic entity as they are and to share only the common internal core AI model of each economic entity.
With the development of natural language processing deep learning models represented by Transformer, and the combination of federated learning methods and newly developed AI-Sharing algorithms, it is becoming possible to create synergy by sharing the core intelligence of economic entities regardless of the language and individual characteristics of global economic entities.
As the possibility of such an AI-Sharing platform is being confirmed, the largest AI engines at each level, such as healthcare, commerce, transportation, finance, smart farms, manufacturing, robots, and smart cities, are expected to be built by AI-Sharing methods. Adopting this method has the great advantage of being able to cooperate without signing a data agreement between economic entities.
AI-Sharing will not just be about sharing artificial intelligence, but a new type of platform that supplies various services based on AI engines developed and maintained in an AI-Sharing manner will emerge in the private and government sectors, as well as in the global business environments. The AI-Sharing platform will enable many individuals, businesses, and institutions to collaborate, which was previously impossible, and AI-Sharing, which will begin on a small scale, will gradually expand, generating a big positive change in the socio-economic system.
Professor Kyoung Jun Lee, School of Business and Big Data Analytics at Kyung Hee University.
Professor Kyoung Jun Lee worked in various academic-industry cooperation projects with Samsung Electronics, LG Electronics, Naver, BC Card, SKT, KT, Shinhan Investment, Busan Bank, Hyundai Motors, and more, and he is currently focusing his research on User-Centric Artificial Intelligence with Harex InfoTech. Professor Lee was a member of the Korean Government 3.0 Committee and the 4th Industrial Revolution Strategy Committee. He has given lectures on internet business, IoT, future technology and consumption revolution, artificial intelligence, business models, and more at EBS Sunday Invitation Special Lecture (2000), Oh My Future (2016), My Job in Moon (2022), at KBS Jang Young Sil Show (2015), at CBS Sebasi Talk (2015), at YTN Science (2018), at MKYU Seven Tech (2021) and many more.
Original Source: Reproduced from the Korean version of the IT Chosun on June 22, 2022
