The third boom in artificial intelligence (AI) has been on the rise. Looking back on the development history of AI so far, the AI systems in the first boom (late 1950s to 1960s), based on simple search technologies, were just like toy systems. Here, we call them "the first generation of AI". In the second boom (1980s), rule-based AI systems which use handcrafted rules and knowledge were developed. Here, we call them "the second generation of AI". In the third boom (2000s to now), machine-learning-based AI systems which find and use rules/models from big data is widely utilized. We call them "the third generation of AI". Especially, deep learning is used as the key technology of the third generation of AI, and various applications of deep learning have been developed. However, several limitations of the third generation of AI came to be clear. In order to overcome them, a new scheme of AI research, which we call "the fourth generation of AI", is required.

This strategic proposal describes and discusses the target direction, important research and development issues, and funding program design strategies for the fourth generation of AI to make Japan internationally competitive in the research area of AI.

The target direction of developing the fourth generation of AI is to realize a new generation of AI that adapts to our society and grows with human by integrating "fast response intelligence" such as deep-learning-based AI and "deep deliberation intelligence" such as knowledge/symbolic reasoning. The reasons why we set this direction are as follows.

Firstly, it is necessary to overcome the problems (limitations) of the third generation of AI. The third generation of AI has three problems: (1) it requires a huge amount of supervised training data and computing power; (2) it cannot handle out-of-learning situations which occur frequently in the real world; (3) it is not good at high-level processing such as semantic understanding and explanation while being good at pattern processing. In other words, the third generation of AI is considered as fast response intelligence which corresponds to perception and motion systems, starting from inductive function.

To solve these problems, a deep deliberation intelligence should also be considered. The feature of deep deliberation intelligence comes from linguistic and logical systems, based on deductive symbolic processing. So, integrating the feature of the second generation of AI with the third generation of AI can be a good approach.

Secondly, it is mandatory to comply with social principles such as "Social Principles of Human-Centric AI" defined by the Japanese government. AI application systems should have some new mechanisms to satisfy the social principles so that they can be reliable for the society and human. Not only they can learn and grow without a huge amount of supervised training data, they should be friendly and adapt to our human-centric society. In the human-centric society, we believe the AI application systems should play a role of competent assistants under human control.

Thirdly, research on human intelligence and human brain can be helpful to the fourth generation of AI research. For example, a human baby can learn language without a huge amount of supervised training data, and also we can apply what we learned in some case to other cases.

In this proposal, from what we have discussed above, we propose the following four important research and development issues which should be tackled toward the target direction.

• 1. Research on the integrated AI which can acquire knowledge and grow without supervised training data.
• 2. Research on the integrated AI which can ensure safety and robustness in the real-world.
• 3. Research on the integrated AI which enables context-aware reasoning and actions.
• 4. Research on the integrated AI which covers both fast response intelligence and deep deliberation intelligence, based on exploration of fundamental mechanisms and principles of human intelligence.

The essential point in these four issues is the integrated AI, that means integration of fast response intelligence such as deep-learning-based AI and deep deliberation intelligence such as knowledge/symbolic reasoning. Based on the integrated AI approach, it is expected to ensure reliability and human-friendly features which are required from our society, as well as to overcome the problems of the third generation of AI.

The above-mentioned issues 1, 2, 3 are challenges to realize functionality, and the issue 4 is a principle quest challenge. We are concerned with that the function realization challenges may only solve their own issue, not bringing the best solution overall. The principle quest challenge is also necessary, then both of the two type challenges should be tackled in parallel. The synergy between the function realization challenges and the principle quest challenge accelerates technology evolution toward optimally integrated AI.

Furthermore, some points to keep in mind in designing funding programs for such research and development include the following.

Firstly, the strategic positioning and focused targeting of research on the fourth generation of AI are important in the current situation where the United States and China are accelerating AI technology development with massive investments and the Japanese government are also promoting AI-related programs based on "AI Strategy 2019". In this point, the proposed research challenges can contribute to the real-world AI and the trusted quality AI, which are enhanced in AI Strategy 2019 and are required in the fourth generation as well as the third generation of AI. Here is a good chance that Japan can precede the world by utilizing advances in computational neuroscience and cognitive developmental robotics.

Secondly, it is important to strengthen research resources such as human resources, collecting data, and computer facilities. As for the human resources, senior leaders in the AI area are now emphasizing the importance of this direction and young researchers' challenge should be encouraged. As for the collecting data, building a large-scale training data set which includes real-world interaction and dialogue is important to accelerate research on for the fourth generation of AI. As it is not started worldwide yet, Japan has a chance to get advantages in new training data building by close collaboration between industry, government and academia in some focused domains (e.g. manufacturing, robot, education, medical). As for the computer facilities, joint use and continuous enhancement of computational resources such as ABCI, Fugaku and Raiden are essential.

Additionally, it is very important to consider the relation of human with the society, ELSI (ethical，legal and social issues), and other non-technological issues. For this purpose, we would claim that we take comprehensive and cross-disciplinary approach to the next AI research, making the best use of not only this proposal "the fourth generation of AI" but also the previously published two proposals: "AI software engineering for reliable and safe AI application system development" and "information science and technology for decision-making and consensus-building".