Widely accessible advanced machine learning will result in expanded use of AI (Artificial Intelligence). AI will increase convenience, resolve intellectual labor shortages, and drastically advance science. Mastery of AI will become a critical component of competitiveness.
In 2016, the big news on Artificial Intelligence (AI) was the development of a technology called deep learning, which mimics the brain cell activities of animals. Deep learning is expected to have higher accuracy than past AI technology and features extraction, which can also be automated, making it available to a wider range of users. In addition, open environments and software are accelerating advancements in deep learning, and TensorFlow, DSSTNE, CNTK and other frameworks and cloud environments for deep learning have been created, further enabling its use. For specific applications such as image recognition, there are now web services that let users perform deep learning and even prediction without programming.
The use of deep learning is rapidly becoming widespread. For example, farmers are using AI to learn images of cucumbers and distinguish ratings, and dermatologists are using AI to learn images of symptoms and diagnose skin cancer. In the future, the use of deep learning is expected to become even easier and commoditization will likely follow.
AI applications are expanding in various fields by leveraging deep learning. AI is already used: in healthcare for diagnosis and drug discovery; in finance for stock trading and credit decisions; and in retail for marketing and customer service. In addition, AI technology supports our daily life in things we take for granted such as web search, path search and translation. Thus, AI will most likely spread to all fields in a natural way.
Robotic Process Automation (RPA), a system that lets digital robots operating with rule engines and AI automate white-collar tasks, is also becoming popular. RPA automates tasks without programming and it is currently used to perform rule-based automation of routine tasks. However, it is believed that advanced AI will be used to automate more sophisticated, non-routine tasks such as analysis and decision making in the future. Some predict*1 that more than 100 million global, full-time intellectual workers will be replaced by AI by 2025. This will likely contribute to the resolution of labor shortages.
AI is also being tried in research applications. For example, AI was able to reproduce the condition of a gas that needs complex control and that is difficult to reproduce. Using a method that would never occur to humans, it completed the task in a matter of an hour. Soon AI may advance rapidly in research-based fields such as science and physics, enabling dramatic progress.
AI is rapidly advancing in image and voice recognition. Moreover, it is still evolving in other areas. One of the evolving areas is the understanding of meaning. In addition to using text to explain the contents of an image, recent studies are actively trying to generate images based on text, thereby generating images that are close to the “meaning” of the text. Inter-exchangeability between language and image may mean that AI is getting close to understanding meaning.
Another area of evolution has to do with the issue that large amounts of teaching data are necessary for AI to learn. Humans have a natural ability to learn inductively based on events that are occurring, and to identify an object after it has been taught only once. For realization to mimic this ability, a technology called deep reinforcement learning is attracting attention. Reinforcement learning is the autonomous learning of the action to take next or the condition that needs to exist based on experience. Deep reinforcement learning combines reinforcement learning and deep learning. With this technology, there is no need to provide teaching data necessary for machine learning in advance. All that is required is to set up the desired actions and conditions. AI will then use repeated trial and error to learn the task. This technology is already used in AlphaGo, an AI Go playing machine that surprised the world by beating a professional Go player in 2016, in self-driving cars and in robots used in factories.
Many studies are underway to make machine learning possible based on small amounts of teaching data for tasks to which reinforcement learning cannot be applied. The fundamental concept of this is called transfer learning, where the knowledge acquired when learning one task is applied to the learning of another, streamlining the learning process. For example, in an image categorization task with limited amounts of data, teaching just one image from a category that AI had not yet learned enabled it to achieve an accuracy that was almost equivalent to a case where large amounts of data were taught. If this technology is implemented, AI is expected to learn things and events at a dramatic speed, furthering expanding its use.
While the development of deep learning is remarkable, it is not versatile. Therefore, it is critical to understand the advantages and disadvantages of deep learning before identifying its application areas. Under some circumstances, rule bases, probability/search models and other traditional algorithms may have to be selected. For example, in the world of chess humans cannot beat AI. On the other hand, in advanced chess collaboration between AI and a human attain better results. As a result, it is critical to identify the areas and the best method for such collaboration.
Even though there are developing technologies such as transfer learning, where small amounts of data enable learning, large amounts of learning data are still necessary to achieve a practical level of accuracy for AI in a new area. For this reason, it is necessary to determine how much learning data can be prepared in advance, and whether or not a system can be built that can perpetually accumulate data during operation based on feedback. In addition, pre-processing such as data cleansing, and the know-how on parameter tuning are important to achieve high accuracy. Machine learning has reached a point where a certain level of accuracy can be achieved. However, the knowledge and technology for utilizing AI will be a critical differentiator for companies in the future.
The high accuracy achieved by deep learning, based on complex network structures similar to actual brain neurons, have become possible thanks to improvements in computer processing. At the same time, the process path to the answer has also become complex, making it difficult for humans to understand the judgments and reason for the end result. This is one of the challenges of deep learning. For instance, while looking at an image of an elephant, humans can recognize it as such, although they are probably not making a judgment based on logical reasoning. However, if they are asked for the reason they have decided the image is an elephant, they site reasons such as its long trunk and huge ears. Thus, similar to humans being able to explain the reason after the fact, AI may be able to provide accurate explanations for the process path to the decision. A project*2 is underway that aims to enable the explanation of the basis of output results. If AI can acquire the ability to explain its judgment and reasoning in the future, it will be easier to improve accuracy and apply it to cases where human lives are involved, such as in self-driving cars.
Discussions are also underway on the singularity of AI’s abilities to exceed those of humans. Many non-profit organizations have been established with the purpose of: keeping AI from becoming private and abused; developing AI to contribute to society; evaluating AI’s impact and establishing development principles*3. Social, ethical and legal issues surrounding AI will need to be solved. These efforts will increase in importance in the future. Although AI is not yet fully developed, it is important to discuss its development and effective use as well as its risks now. Doing so will contribute to the benefits of using AI for humans, such as the resolution of labor shortages and energy issues.
*1 McKinsey Global Institute – “Disruptive technologies: Advances that will transform life, business and the global economy.” May 2013 http://www.mckinsey.com/business-functions/digital-mckinsey/our-insights/disruptive-technologies
*2 Explainable Artificial Intelligence http://www.darpa.mil/program/explainable-artificial-intelligence
*3 Open AI http://www.openai.com/ Partnership on AI http://www.partnershiponai.org AI Network Social Promotion Conference http://www.soumu.go.jp/main_sosiki/kenkyu/ai_network/