We are now in the third Artificial Intelligence (AI) boom. What is driving this boom is "Deep Learning" and other forms of machine learning technology. Machine learning technology does not only excite the research community and print media. It is also embedded into various systems and rapidly put into practical use in the real world.

Machine learning technology inductively defines a system's operation from vast amounts of data, and thus it differs greatly from the conventional development style whereby a program code is written and then the system is defined accordingly. Therefore, the methodologies and technologies that have been built and developed for conventional software engineering are not always applicable. This makes a new way of thinking necessary for the definition of system requirements, performance guarantee, and quality assurance. Hence a paradigm shift is occurring in the system development.

With this paradigm shift, the skills and methodologies required for system development need to be renewed, and the software industry and system integrators may lose its competitiveness if they fail to keep up with this shift. Moreover, if systems embedded with machine learning technology rapidly appear in society, even as approaches to performance guarantee and quality assurance remain undeveloped, the resulting difficulties and accidents could produce unexpected social problems.

The paradigm shift in system development that is described here could have major implications for the industrial world. However, in comparison to R&D on AI and machine learning technology themselves, R&D to methodologies for developing the systems remains insufficient.

In light of this, we, through this proposal, aim to establish "AI software engineering" and recommend the promotion and strengthening of R&D for this purpose.

In the case of conventional system development, technologies and methodologies for maintaining safety and reliability and developing systems efficiently have been established within software engineering. On the other hand, AI software engineering could be described as "next-generation software engineering." It targets not only conventional systems but also systems that apply AI, including machine learning-type components, and refers to technologies and methodologies for maintaining safety and reliability in them.

It should be noted that, while we refer to this kind of next-generation software engineering as "AI software engineering" in this proposal, it roughly corresponds to what is known as "machine learning systems engineering" in Japan and as "Software 2.0" in other countries.

We view R&D themes for developing and establishing AI software engineering as a new academic field and for Japan's attainment of international competitiveness in this technology from two aspects. The first is the "systematization of AI software engineering" and the second is "important technical challenges as basic research."

First, system safety and reliability cannot be achieved with just one technology; they must be attained through multifaceted and comprehensive initiatives. They must be tackled with clear vision and good balance and with a grasp of the whole picture. Thus, we put forth "the systematization of AI software engineering" as a research theme in the sense of preparing a technical system with a panoramic view and in parallel with the development of individual technologies.

Additionally, multifaceted and comprehensive initiatives to AI software engineering could make the following approaches: involving the systematization of knowledge gained from practical application ("Approach A"), and an approach involving academic study concerning theory and principles ("Approach B"). Of these, Approach A will certainly take place naturally, particularly in industry. On the other hand, Approach B is essential as a more fundamental way of dealing with the paradigm shift in system development.

Our belief that, rather than leaving things to Approach A, which is primarily focused on industry, the national government should also put priority on promoting Approach B for academic research. This the reason why we put forth "important technical challenges as basic research" as the second R&D theme. Here, we present the following four challenges as being particularly important: (1) quality assurance in machine learning in and of itself; (2) ensuring safety in the entire system; (3) tackling the "black box" problem; and (4) design of an engineering framework for efficiently resolving problems.

Furthermore, in addition to basic research topics, the government should simultaneously study the formulation of safety standards (and mechanisms for thirdparty evaluation) and intellectual property rights concerning the reuse of trained models.

New management measures must be implemented to promote this kind of R&D in AI software engineering. The following policies and approaches will be the key to designing and executing effective measures.

• Top-down acceleration of problem-solving-oriented basic research
• Promotion of the technical fusing of AI and software engineering, creation of new technical fields, and reinforcement of business competitiveness
• Promotion based on the close linkage of three activities; namely, (1) academic research and human resources development, (2) technical validation in practical application, and (3) standardization
• Advance validation in the emerging competitive field of automated driving and expansion to other fields
• Prompt promotion of measures that take advantage of Japan's existing superiority

As was stated above, at a time when AI and machine learning technologies are becoming embedded in the various systems of society, the promotion and establishment of AI software engineering are essential for securing the safety and reliability of those systems and for raising Japan's international competitiveness by using the quality of its AI and machine learning as an advantage.