The possibility of bi-directional conversation between text and images in machines makes us realize that the machine is now able to understand the language.
There has been a utility surge in a new technology called Deep Learning that imitates nerve cell of living organisms. Lots of businesses have already started using this technology that includes for diagnosis and drug discovery in the medical field, for stock trading and credit judgment in the financial field and for marketing, customer service in the retail field etc.
Deep Learning has gained popularity because it is not only more accurate than conventional AI technology, but it is also possible to automate the features that can extract the necessary information for learning and is also very user friendly with any layman. The popularity of Deep Learning increases by the day because the framework or cloud environment for implementing Deep Learning such as TensorFlow, DSSTNE, Chainer have been provided. Applications like image recognition, web services also allow Deep Learning to be carried out using a mouse (without programming) to execute predictions (※1).
Due to the commoditization of Deep Learning the services currently using AI will become so habitual that maybe in a few years’ time no one will be aware that they are using AI at all. AI is anyways now being used in all areas.
Evolving AI technology
AI is rapidly developing focus on image/ speech recognition and is bound to evolve further and there is a lot of awareness already...
In recent years, there has been a surge in not only describing images with text but also of generating images from text, and the machines can generate images close to the “meaning” of text. The possibility of bi-directional conversation between text and images in machines makes us realize that the machine is now able to understand the language.
We also have a lot of progress towards problems that require a large amount of teacher data for learning. Even though we cannot explicitly give a correct answer, human beings can learn from nature from what is happening, and can discriminate objects by being taught once. In order to realize this capability, a technique called deep reinforcement learning is gaining attention (which combines deep learning with reinforcement learning) which autonomously learns from experience, the behavior and the state in which it should exist.
In this technique, it is not necessary to give training data for machine learning in advance. By setting the target behavior and state, the machine by itself will learn on its own by repeating trial and error. It has already been used in “AlphaGo” of Go AI which won the professional shogi player in 2016 and surprised the world.It also optimized the cooling facility setting in the data center, etc.
Tasks in which reinforcement learning cannot be applied, there are now efforts to realize learning with less teacher data. In image classification with limited data, by simply learning one image of unlearned category, it is now possible to classify images with almost the same accuracy as learning a large amount of data.
It will not be long before AI gains the generic ability to identify by just watching it once. If this is realized, it will be possible to learn these things and events around the world at a dramatic speed and the use of AI is expected to expand remarkably.
The precision realized by deep learning is a major factor that enabled a very complicated calculation that improved the processing capacity of the computer. On the other hand, the process to reach the answer becomes complicated, and it becomes very difficult to explain this process to the people.
Although humans have the ability to recognize an elephant when we see an image, we do not need to think logically the reason for our judgment. However, when asked as to why we judged it as an elephant, we have reasons to believe because we know that it has a long nose and big ears. Using the above explanations or by analyzing the output of internal processing of AI, it may be possible to explain exactly the process of how to reach the judgment.
A project has been launched to explain the clue of the output judgment(※2), and if AI can acquire the ability to explain the reason for judgment in the future, it will be easy for us to improve the accuracy and to apply AI to situations involving human life for e.g. autonomous driving.