AI-Sharing is the business term for Federated Learning. Federated Learning was first introduced by Google in 2015.
Here is how it works. Instead of sending data from user terminals such as smartphones to servers, the server sends AI models to smartphones. The AI model learns using the data existing in an individual's terminal, and only the learned parameters (weights in the case of deep learning) are sent from the terminal to the server. The server combines parameters transmitted from multiple terminals (mathematically weighted average) to create one model. Then the server sends that created model back to the user's terminal and shares it.
This eliminates the need for large corporations like Google to take users' personal data. By sharing only AI among users, it is possible to create an AI model with similar performance as sharing data directly. In this way, privacy can be protected, and the use of computers for learning can be distributed as well. Since only parameters are shared, not data, there is also an effect of saving communication costs.
This AI-Sharing model is possible between companies and users, and between companies and companies. It will still be necessary for some companies to provide their own AI services, but by utilizing AI-Sharing methodology, it is possible to create more advanced AI engines and services that suit each company’s characteristics by sharing intelligence and knowledge with each other.
AI-sharing is most active in the medical field. EXAM is the first AI-sharing experiment related to COVID-19. It was reported in Nature Medicine in October 2021. 20 medical institutions in four continents around the world did not share data on COVID-19 patients but learned through AI-Sharing.
In the EXAM sharing experiment, AI was trained with data from each institution without centralizing the data. As a result, it was confirmed that with the EXAM sharing experiment, AI works better and also helps develop global models that can be more generalized, than the 20 models developed separately by each hospital.
In particular, it showed a sufficient performance even when only 25% (weighted) was shared, in a situation where each hospital did not share all of the learned AI models.
Smaller hospitals, especially those with less data, benefited more, and the largest hospitals also benefited from AI-Sharing. It has been shown that the polarization that can be caused by digital and artificial intelligence can be resolved through this method.
In addition to EXAM, cases of UCADI (Unified CT-COVID AI Diagnostic Initiative) and PriMIA projects were reported to Nature Machine Intelligence in 2021. Here, the global AI-Sharing model had better performance in every category over the AI model developed by each hospital. Even if each hospital did not share its data with other hospitals, it was able to perform better with AI-Sharing. The CAreFL project, which studied the methodology of preventing opportunistic attitudes that may appear in the process of sharing AI models and fairly evaluating contributions, won the American Society for Artificial Intelligence's innovative AI application award in 2022.
Artificial intelligence sharing does not only apply to the medical field. Various cases such as AI-Sharing among mobility service companies to predict traffic flow, AI-Sharing among credit card companies to prevent financial fraud, AI-Sharing among welding robots at smart factories, and AI-Sharing for personal health monitoring are being announced in the 2020s.
The power of AI-sharing has not yet been well transmitted to academia and industry. However, pioneers are rapidly commercializing. In the medical field, non-profit organizations such as MedPerf (Open Benchmark Platform for the Medical World) were established. Related startups have been established in advanced AI countries such as the United States, Canada, and Germany. It is similar to the situation where several web agencies and startups were developed after the advent of web technology.
At first, there will be many AI-sharing agencies that develop solutions and services for AI-Sharing in existing business groups or specific applications. However, in the end, it is expected that a global digital platform company based on AI-Sharing will emerge. In Korea, Harex InfoTech Inc., which has long pioneered sharing platforms and platform sharing business models by advocating User-Centric Artificial Intelligence, is recently introducing the AI-Sharing platform business model to the world.
Professor Lee from Kyung Hee University (School of Business & Big Data Applications)
Professor Kyoung Jun Lee graduated from KAIST's Management Science with an undergraduate, Master’s, and Doctor’s degree, and graduate from Graduate School of Public Administration, Seoul national university with Mater’s and Doctor’s degree. He also studied at Carnegie Mellon, MIT, and UC Berkeley. In 1995, he was the first Korean to win the Innovative Artificial Intelligence Application Award by the American Society of Artificial Intelligence (AAAI) and also won the 1997 and 2020 (best awards). He published three research papers in AAAI's quarterly journal AI Magazine. He is currently focusing on research of User-Centric AI with Harex InfoTech. He served as the chairman of the Korea Intelligent Information System Association (2017) and is currently serving as the vice-chairman of the Leading Future Agendas of Business & Society and the senior vice president of the Korea Society of Management Information Systems. He selected 25 Korean AI startups in 2020 and 100 Korean AI startups in 2021. In 2018, he was awarded the Presidential Citation for e-government merit and is currently a general member of the National Academy of Engineering of Korea
Original Source: Reproduced from the Korean version of the IT Chosun on June 21, 2022
