Asilla, a developer of security systems with world-class behavior recognition AI at its core, launched the Human Science AI Research Team (HSAR) in February 2023 to further improve the accuracy of its AI. In this article, we interviewed Mr. Masahiro Wakasa, CTO of Asilla, the founder of HSAR, about his research efforts on "Generative AI" and its future development.
"Generative AI", as the name implies, refers to "AI that generates data based on certain inputs". For example, it is an AI that generates images, text, music, etc., through its own learning based on past data or human instructions. It is truly a creative AI. Generative AI has been rapidly advancing and further research has been conducted since the announcement of the image-generating GAN in 2014. AI has become a hot topic in Japan as well. Generative AI has matured technologically and is now in the process of being put to practical use in multiple fields. We are now in the phase of "how to connect AI-generated products to new value". We are now at the stage of using this technology in terms of behavior recognition AI.
The HSAR is an attempt to have a generative AI automatically generate behavioral data based on human behavior. By behavioral data, we mean data related to human behavior, such as, for example, that a person is "fighting" or "has fallen down and stopped moving. We are conducting research and development of AI based on such behavioral data on a daily basis, but in the past, data acquisition was done in a way that had a strong analog component, in which we actually acted and recorded our actions. The improvement of AI accuracy is proportional to the amount of data to be learned. We are now considering the possibility of using generative AI to collect more data, and to generate new behavioral data based on the vast amount of data accumulated in the past. Then, the AI will use the generated data to learn again. We are currently conducting research to realize such a virtuous cycle.
It is very difficult to determine how to correctly identify the data that "the human behavior data generated by AI is really human-like". After all, computer science techniques alone cannot answer the question, "Is this behavioral data human-like? Or is it a mechanical movement? I feel that it is very difficult to properly discern whether this behavioral data is truly human or mechanical. I believe that human science research is necessary for this determination. Since behavioral data is the foundation of AI research, utilizing only truly human-like behavioral data will further improve the accuracy of AI's prediction models, which will lead to new learning. We hope to see a cycle in which an AI that can truly infer human behavior is created. Although we are still in the research phase, we are beginning to see some very good results. We have been steadily accumulating data on human behavior for many years before we began to use generative AI. It is precisely because we have a vast amount of human-like behavioral data that we feel we are entering the next phase of our research.
We have not strictly quantified the data explicitly, but we have dramatically improved the speed of data generation. For example, let's say you have to create more than 10,000 pieces of data for a single action in order to correctly detect it. With conventional data acquisition methods, this can take up to two weeks. On the other hand, with generative AI, data generation can be completed in as little as 1-2 hours. Moreover, the conventional method requires a person to be involved in the process, which requires labor costs, whereas with generation AI, all you have to do is press a keyboard button and wait. Therefore, no human labor is required for data generation. We feel that this not only dramatically improves the speed of data generation, but also reduces costs.
The use of generative AI is a turning point in our future research. The HSA research team will continue to study the use of generative AI to further improve the accuracy of AI. However, the data to be utilized must be of high quality as well as quantity. Therefore, data acquisition is time-consuming and costly. Ethical issues surrounding cheap labor have also recently become a problem in the area of data acquisition. If AI is able to produce correct data on its own, we believe that this will help resolve these ethical issues.
Generative AI, which has the characteristic of "AI creates itself," has evolved remarkably over the past year or two. I have the impression that most generative AI is currently focused on the entertainment domain, such as the generation of beautiful images and illustrations. This trend is expected to continue in the entertainment domain, and at the same time, it is also expected to expand to solve problems in the business domain. One of these problem-solving areas is the area of "behavior-aware AI," which is one of our strengths. As we envision such a future, our entire team at Asilla is working together to advance research and development toward that future.
After completing his graduate studies at Tokyo Institute of Technology in 2018, he joined JGC Corporation. Kuwait, where he was stationed at construction sites and engaged in plant design IT work. He then joined Asilla Corporation in 2020, where he was in charge of product development related to action recognition AI, and was appointed to oversee the IVA Solutions Division, becoming Executive Officer CTO in April 2022 and Director CTO in March 2023.
