Research

AI's Current Insufficient Reliability

There has been a rapid growing of interest in Artificial Intelligence (AI) recently. Generative AI systems such as ChatGPT create texts or images in response to prompts, but these systems often produce ridiculous responses including biased and false information. At its current level, AI certainly cannot be said to be "trustworthy."

In terms of trends in AI research, there had been much research into symbolic AI, where knowledge is expressed as symbols, but the dominant trend in AI has shifted to deep learning based on neural networks. In the course of AI research history, the two distinct fields have been separated: Knowledge Representation and Reasoning (KRR), which deals with the symbolic world, and Machine Learning (ML), which deals with numerical/vector spaces.

Inoue has been involved in AI research since the 1980s from a computer science perspective. In his view, the current situation, where AI produces results that are clearly inappropriate from a human point of view, needs to be rectified as soon as possible. This means improving the reliability of AI; but as well as ensuring that it will not produce inappropriate or inconsistent responses, we need to think about other factors, such as making systems robust against unexpected changes and noise, in order to make AI more reliable.

Significance of Integrating Machine Learning (ML) and Knowledge Representation and Reasoning (KRR)

Inoue believes that the reliability of AI can be improved by bringing together the two areas of research that have become separated: KRR and ML. His current research can be summarized as "providing a basis to handle the whole intelligence process of recognition, learning and inference on a common mathematical ground." What he means by a "common ground" is an algebraic technique using algebraic data structures such as vectors, matrices and tensors to represent logical formulas and constraints.

Reasoning problems are usually solved using symbolic processing and algorithms, but by replacing them with numerical calculations, we can make the problems more robust and scalable. Inoue's algebraic technique belongs to the field of neurosymbolic AI, which deals with both symbols and sub-symbols, but is based on an original approach where logic programs are represented by tensors.

Inoue often has opportunities to present this research at international conferences and has been invited to give talks at three international conferences in FY2024. As well as publishing research results in papers, the data and programs developed by the research team are made openly available. As previously unsolved problems are gradually being solved, his research results gain more and more attention.

Aiming for Trustworthy AI Through International Cooperation

One technology that Inoue is currently keen to develop is "AI incorporating constraints." Here, "constraints" refers to being expressed as symbolic knowledge. This research is looking at incorporating constraint knowledge into machine learning and generative AI, and how far this could improve reliability.

One example of the results of this research is seen in ROAD-R Challenge for NeurIPS 2023: the Road Event Detection with Requirements, where participants competed to create an object detection system for autonomous driving, whose predictions are compliant with the requirements. Inoue's team achieved the first and third places in the two respective tasks. In this AI system, logical constraints have been added as requirements to object recognition technology based on machine learning, so this is an example of Inoue's research theme of integrating KRR and ML. Applying a similar idea to generative AI should help to make generative AI more reliable.

Some people claim that Japan is falling behind other countries in terms of AI research. In the course of his career, Inoue has conducted joint research with overseas research institutes and researchers, and he intends to push ahead with his own research through international collaborations in pursuit of reliable AI. He believes that in this way, Japanese AI research can be recognized as world-level.

Although AI may still be immature in many respects, there are more and more examples of AI being implemented as a function of various systems, which are of great benefit to us. The evolution of AI is sure to continue, so we would like people to keep a close eye on researchers working on fundamental AI research, like Inoue, without being swayed by the temporal AI boom.

PDF Download