Converting the Tacit Knowledge of Experts into Explicit KnowledgeFeatured Technologies
Behavioral Imitation Learning Technology
October 19, 2022
With the decreasing birthrate and aging population, staffing is now becoming a critical issue in many industries. How should the skills of retiring experts be passed on? How can newly hired staff master and share skills in an efficient manner? NEC's "Behavioral Imitation Learning Technology" helps to resolve such issues. But what exactly is this technology? We spoke with a researcher about the details.
Learns from past data to recommend the measure which is optimal for the current situation
― What kind of technology is behavioral imitation learning?
It is a technology in which AI learns the tacit knowledge of experts and imitates those skills and behaviors. For example, if it learns the treatments in a medical setting and data about the progression of patient symptoms, it can imitate the skills of medical personnel with extensive experience. If you ask the trained AI about treatments corresponding to the current symptoms, it can compare them with past data to propose the optimal method of treatment. It is a technology which is effective in visualizing tacit knowledge and sharing generalized skills.
Currently, there is a growing shortage of staff in various industries. Passing on skills is a critical issue particularly in industries where experts continue to retire due to the population decrease and aging population. In order to address this problem, NEC focused on "imitation learning" technology, which has primarily been used in the robotics field, and advanced its research and development.
Imitation learning differs from "reinforcement learning," where the AI learns on its own from repeated trial and error, in that it inputs past example data to learn from it. Therefore, you must collect a large volume of successful examples to produce a satisfactory level of accuracy, and this point was a bottleneck for implementation. That is even more true when it is applied to human skills rather than the traditionally utilized robotics.
In response, NEC developed a system that can learn not only from successful examples but also examples of failure which could not be utilized with conventional methods. As a result, this made it possible to increase the AI accuracy in a complementary manner based on both the successes and failures, and we succeeded in producing a sufficient level of accuracy at actual sites as well.
Rapidly creating a high accuracy AI
― Specifically, what kind of system is this?
A simulator is used to train the AI, but at this time it combines three different types of AI which consist of "AI that imitates successful examples," "AI which identifies successful examples," and "AI which identifies examples of failures." These AIs apply the "GAN (Generative Adversarial Network)" framework which is primarily used in the fields of image generation, etc. In a GAN, the AI is trained to be able to create imitation data with a higher accuracy by combining two AIs consisting of an "AI (Generator) that imitates and creates data" and an "AI (Discriminator) that identifies whether it is the real thing." By having both AIs antagonistically compete with each other, the Generator develops into an advanced AI that can deceive the Discriminator.
In behavioral imitation learning, it utilizes an "AI that imitates successful examples (Generator)" and an "AI that identifies whether it is a successful example (Discriminator)" while also introducing one more Discriminator which is an "AI that identifies whether it is an example of failure." The AI which imitates and creates successful examples cooperates with the AI that tries to identify examples of failures to produce high accuracy imitation data (= imitated successful example) that will avoid being judged as an example of failure to a greater degree. By approaching from both success and failure in this way, we successfully resolved the problems of the volume of data and accuracy required for learning.
― Are there any other features?
Another major feature is that it is able to capture behavior patterns "not as points but as a line." If you focus only on imitating behavior, then it would hold true even if you only think about that situation (point). That is because it would be suffice to say that in this situation, statistically speaking the probability of this kind of behavior is high. However, you do not know if you can achieve the ultimate goal with that alone. For example, when a physician thinks about an action to lower the patient's blood pressure, naturally the blood pressure would lower if many hypotensive agents were administered. However, that would, of course, likely harm the health of the patient. Ultimately, the goal is to improve the state of health of the patient. For example, even if medicine is administered and the patient's condition improves for short period of time, it would be meaningless if the patient's health subsequently deteriorated. Therefore, it is important to capture and approach the behavior data not as simple dots but as a continuous line. Our behavioral imitation learning is built on this kind of perspective.
Moreover, one more notable feature is that it handles numerical values, text, and other lightweight forms of data. I think that there are various approaches for passing on skills, but our behavioral imitation learning technology does not require the installation of cameras, video data, or image processing, etc. You can directly use numerical values, text, and other data that you already have. Therefore, the introduction and data learning are both extremely fast.
Considering demonstrations in the areas of healthcare, large-scale plants, and sales
― What level of completion is the technology currently at?
We are currently advancing demonstrations focusing on medical institutions. For example, in a demonstration made possible with the cooperation of the Kitahara Hospital Group, we created a smartphone app which supports the formulation of rehabilitation plans. Over a roughly two year demonstration, the Kitahara Hospital Group commented that the validity improved by 46% when creating plans. Moreover, they also mentioned that providing this application to new employees deepened their understanding and thinking, which led to sharper and more precise questions to leaders and their senior colleagues.
In addition, we also collaborated with the Tohoku University School of Medicine to advance research into a technology that recommends the dosage to administer to diabetes patients. Controlling the blood glucose level is an extremely delicate medical practice that can only be performed by expert medical specialists. We thought that perhaps AI could be applied in this case and have been working on verifying that application. Currently, we have obtained results that it can provide numerical values with a fairly high accuracy.
In addition, we are also exploring the application of this technology to the operation of plants and factories where a shortage of staff is becoming critical due to the retirement of experts. As a slightly curious example, we are also considering applying it to the sales domain with the goal of improving sales across the entire sales force by learning the skills of top salespeople.
Behavioral imitation learning is a technology with the potential to be utilized in various domains. In addition, we would also like to advance the application of this technology with the goal of social implementation in a wide range of sites while refining the algorithm.
- ※The information posted on this page is the information at the time of publication.
Behavioral imitation learning was developed by NEC based on technology that is typically called "imitation learning." It is a technology which is primarily used in the robotics field, but NEC focused on applications of the technology to learning the tacit knowledge possessed by humans and continued to conduct research. Moreover, the realization of an improvement in AI accuracy by training the system not only on successful examples but also examples of failure shows NEC's originality.