OMRON Develops Decentralized Learning Technology, "Decentralized X," Enhancing Performance of AI without Aggregating Data- Allowing multiple machine learning models to integrate themselves for expanded social implementation of AI -

• November 13, 2019

OMRON Corporation (HQ: Shimogyo-ku, Kyoto, President and CEO: Yoshihito Yamada, hereinafter, "OMRON") announced today the development of a decentralized learning technology that enhances the performance of artificial intelligence (AI) by integrating multiple machine learning models*₁ rather than aggregating the data needed for AI learning in one place. Dubbed as Decentralized X (patent applied for*₂), this innovative technology allows businesses without access to massive pools of data to introduce high-performing AI, thus assisting small- and medium-sized enterprises (SMEs) to solve social problems without the need for big data access.

To improve the performance of AI, it is necessary to gather together a sufficiently large amount of data in one place and have the AI learn from it. However, this gathering process incurs a huge cost and use of numerous human-hours, as data is located in various places. Protection of trade secrets and privacy as well as other restrictions add to the difficulty, thus making it nearly impossible to gather a sufficient amount of data. It follows, then, that collecting data of the required quality in the required quantity poses a major challenge for commercial applications of high-powered AI.

Decentralized X, OMRON's new offering, boosts AI performance by integrating AI machine learning models that have learned in a myriad of settings. Integrating different machine learning models is a technically challenging process, as the structural variety of said models makes it difficult to combine them without compromising their properties. Decentralized X uses a unique algorithm capable of integrating different machine learning models to develop AI, the performance of which compares favorably with that of AI based on learning that aggregates all the data.

For example, with Decentralized X, SMEs without a huge platform for gathering big data can facilitate AI-enabled solutions to day-to-day issues by contributing each of their machine learning models to integrate them for improved performance of their own AI. Even in cases where it is difficult to share data with other businesses due to the confidentiality of data (e.g., manufacturers' quality data and vital data from medical institutions), Decentralized X helps achieve collaboration in a way that enhances the AI performance of each business.

Decentralized X is the brainchild of OMRON SINICX Corporation (hereinafter, "OSX"), an OMRON subsidiary that creates "near-future designs"*₃ for solving social issues through backcasting. Having assembled a team of exceptional researchers from a broad range of cutting-edge technological fields, such as AI, robotics, IoT (Internet of Things), and sensing. OSX is engaging in the research and development of innovative technologies they hope will solve social issues in the near future. Going forward, OMRON will apply Decentralized X to AI currently used in the Industrial Automation Business and other focus domains, thus accelerating the evolution of OMRON's AI, while at the same time encouraging open innovation among different players and social implementation of AI in various fields to work on solutions to social issues.

How Decentralized X works

Main features of Decentralized X

1. Builds AI comparable in performance to AI that has learned from aggregated data, even under circumstances where individual players each own a small amount of data and thus data is not centralized

• ・Provides an environment where businesses that were previously unable to fully benefit from the introduction of AI due to insufficient data find it easier to utilize AI
• ・Reduces communication costs by inviting players to bring together their machine learning models, rather than data

2. Learns from decentralized data since it is open to any AI architecture

• ・Since not all the parameters of machine learning models need to be shared, it learns from the AI of any architecture; thus each AI machine can retain its own characteristics (e.g., a light-weight model that prioritizes processing speed)

3. Enhances confidentiality of data since individual players do not have to disclose their data to each other to have their AI learn

• ・Builds AI while maintaining a high level of confidentiality by combining this technology with security and privacy protection technologies when handling data that must be kept from other players

*1 An algorithm with a specific function, created from data that AI has learned

*2 A patent has been applied for in Japan (current as of November 2019).

*3 A concrete business architecture tailored to solving social issues envisioned by OMRON as it relies on technical innovation to integrate "business models," "technology strategies," and "intellectual property strategies"