This report focuses on two trends in the research and development (R&D) of artificial intelligence (AI) technology, namely "Fourth-Generation AI" and "Trustworthy AI," and provides strategic proposals for R&D to enhance the social value of AI technology and strengthen Japan's international competitiveness, as well as detailed trends in R&D areas closely related to them. A similar report titled "New Trends in AI Research: Japan's Winning Strategies" was published two years ago (June 2021). Since then, generative AI based on foundation models has made great progress and had a significant impact on society. R&D strategies and policies related to them are also under active consideration. Therefore, we have decided to update the report of two years ago and publish it under the title "New Trends in Artificial Intelligence Research (2) The Impact of Foundation Models and Generative AI".

This report consists of two parts: Part 1 summarizes the two R&D trends and strategic proposals based on the impact of foundation models and generative AI, and Part describes detailed trends and international comparisons in R&D areas that are closely related to them.

Part 1: R&D Trends and Strategic Proposals

Generative AI based on foundation models created by ultra-large-scale deep learning has demonstrated extremely natural dialogue response performance and high versatility and multimodality. It is rapidly transforming human intellectual work in general. Its impact has also influenced the direction of "Fourth-Generation AI" and created new challenges in terms of "Trustworthy AI". Part 1 outlines the strategic proposals that we have published so far along the two trends of "Fourth-Generation AI" and "Trustworthy AI," and presents new issues on strategic proposals brought about by the foundation models and generative AI.

The first trend, "Fourth-Generation AI," is the R&D of new AI basic principles and architectures that overcome the problems of third-generation AI, mainly deep learning, in line with the trend of improving the accuracy and performance of AI technology. The deep learning has shown high accuracy and performance, and has a wide range of applications. However, it has been pointed out that (a) a large amount of training data and computational resources are required for learning, (b) it is vulnerable to situations outside the scope of learning and cannot respond flexibly to real-world situations, and (c) it has advantages in pattern processing, but not in higher-order processing such as semantic processing and explanation. The "Fourth-Generation AI" is a direction of R&D that aims to overcome the above problems (a), (b), and (c) by fusing pattern processing using deep learning with language and symbol processing including knowledge and symbolic reasoning in a unified way.

To address problems (a), (b), and (c), humans can learn and grow without large amounts of labelled training data, and can combine what they have learned and apply it to different situations and circumstances. Moreover, it is known that the human brain consumes only about 20 watts of power. Therefore, it is considered effective to learn from the research results and knowledge on the information processing mechanism of human intelligence, and the dual-process model (System 1 + System 2) and the developmental/emergent model (prediction error minimization principle) are being examined. On the other hand, the foundation model (positioned as 3.5th generation AI) that utilizes the ultra-large-scale deep learning has increased accuracy and versatility, and the possibility of intelligence that is far from human has been demonstrated in a sense. However, although some of the above problems (a), (b), and (c) were improved, problems remained in logical inference, logic construction, and real-world operations, and the problems requiring extremely large computational resources became even more serious. Moreover, the mechanism of why the foundation models shows such high performance is not clear. Given this situation, the basic principles and architecture of "Fourth-Generation AI" will be discovered through exploration, fusion, and development of both approaches: learning from the mechanisms of human intelligence and clarifying the mechanisms of foundation models that are far from humans.

The second trend, "Trustworthy AI," is a type of R&D that is different from that focused on accuracy and performance, and has emerged as the relationship between AI and society has expanded. With the expansion and spread of AI technology, the black box problem of AI (the demand for interpretability and explanation), the bias problem (the demand for fairness), the vulnerability problem (the demand for robustness and safety), the quality assurance problem (the demand for quality assurance and reliability), and the fake problem (the demand for the credibility of information) have become apparent. AI social principles and AI ethical guidelines that formalize these social demands for AI were discussed and established at the national and international level. Then the effort to realize and put it into practice is the R&D of "Trustworthy AI."

"AI Software Engineering," aiming to establish and improve methodologies and technologies for "Trustworthy AI," has emerged as a new field of R&D, and is being actively worked on. Furthermore, the development of generative AI has caused not only problems of abuse such as deepfakes, but also a backlash from experts due to the ease of creating texts, images, videos and so on that appear as if they were created by human experts, and concerns about authenticity judgment and credibility, including spoofing. In the face of such uncertainty about truth and falsehood, human "decision-making and consensus building support" and "trust formation in the digital society" have become important issues.

Furthermore, the development of AI technology, appearing as two trends, promotes the transformation of society, industry, and science. Focusing particularly on "AI- and Data-driven Science," which determines the international competitiveness of various industries and science and technology, there has been active effort to enable scientific discoveries that go beyond the cognitive limits and biases of human beings through large-scale, comprehensive hypothesis generation and exploration using AI, as well as to increase the throughput of hypothesis evaluation and verification using robots. The generative AI is expanding and accelerating this potential.

The United States and China are said to be the two biggest players in AI technology development, and it will not be easy for Japan to break into their ranks. However, the AI technology is indispensable in terms of economic security, as it determines the international competitiveness of a country's industry and science. Rather than relying entirely on the adoption and utilization of technologies from overseas, Japan should find points where it can establish its superiority and promote R&D strategically. This report discusses the possibilities from the perspectives of "Fourth-Generation AI," "Trustworthy AI," and "AI- and Data-driven Science."

Japan has declared "Trusted Quality AI" in its AI strategy, and began its efforts both in the research community and industry at an early stage internationally. It is also doing well in international standardization activities, and in response to Europe's movement to tighten AI regulations led by ideas, it is making efforts to build up practical trust toward the realization of "Trustworthy AI." In addition, while the foundation model and generative AI (3.5th generation AI) pioneered by US Big Tech companies is having a major impact on society, basic research leading to "Fourth-Generation AI" is also underway, including the fusion of computational neuroscience and AI, and the cognitive development and symbol emergence robotics, which are research fields originating in Japan. Furthermore, in the field of "AI- and Data-driven Science," Japan proposed a grand challenge called the Nobel Turing Challenge, and efforts in this area are also being made through international collaboration. We believe that such efforts can serve as a foothold for fostering Japan's strengths.

This report also presents an overall view of the R&D issues in the foundation models and generative AI, which are being actively discussed and designed in terms of policies and strategies. The report then discusses the direction of linking the three activities: promoting the use of the foundation models and generative AI, promoting the basic research for next-generation AI models, and implementing and operating the foundation models and their peripheral technologies that support them.

Part 2: Detailed Overview of Notable R&D Areas

Part 2 of the report first highlight the following nine R&D areas closely related to "Fourth-Generation AI" and "Trustworthy AI" (i.e., (1) to (9)). It also introduces R&D trends, technology topics, important future issues, and international comparisons in each R&D area.

Although the third-generation AI to date has demonstrated performance superior to human beings in various applications, the aforementioned problems (a), (b), and (c) have been pointed out. The foundation models and generative AI (3.5th generation AI), an expanded version of third-generation AI, has improved upon some of the issues and has shown high versatility, but as a further development, the R&D of "Fourth-Generation AI" by combining "(1) Perception and Motion AI Technologies" such as image and video recognition and motion control and "(2) Language and Knowledge AI Technologies" such as natural language processing has begun to progress. In addition to technologies development such as deep learning, deep reinforcement learning, deep generative models, self-supervised learning, and ultra-large-scale foundation models, insights into human intelligence gained from research in "(7) Computational Neuroscience" and "(8) Cognitive Developmental Robotics" play important roles in the R&D of "Fourth Generation AI." The relationships between such AI and humans or multiple AIs are expanding due to "(3) Agent Technology."

On the other hand, the AI technologies have spread throughout society, and from the perspective of "(9) AI in Society," AI social principles and AI ethical guidelines, including safety, reliability, fairness, interpretability, and transparency, have been formulated at the national and global levels, and the R&D for "Trustworthy AI" has become an important research issue. Specifically, progress is being made in "(4) AI Software Engineering" to develop AI application systems that are expected to satisfy the above principles and guidelines, "(5) Human-AI Collaboration and Decision-making Support" which aims for humans to collaborate with AI to make better decisions and achieve goals, and "(6) AI- and Data-driven Problem Solving" which leads to the transformation of society, industry, and science using AI and big data technologies.

In addition, the following four R&D areas that are closely related to the above nine areas are introduced: "(10) causal inference," "(11) mathematics of decision-making and optimization," "(12) computing architecture that supports AI," and "(13) trust in society."