Author: Doctor of Technical Sciences, Artezio consultant, Professor Vladimir Krylov.
Artificial intelligence has long ceased to be the technology of the future. Already today, AI is used in many areas – from healthcare and finance to defense and autopilot vehicles. However, if you look into the future, you can see that AI will not only confidently take its place in human lives, but will also become the key technology on which all the most effective projects will be based.
How and why AI should be taught?
When creating modern AI, three basic methodologies of teaching it the required functions are used: learning with a teacher, a self-study, and reinforcement learning. Usually, for each specific task, effective AI is obtained based on one of the mentioned methodologies. In those areas where a person has full understanding of how to perform certain actions correctly and a large number of examples of right solutions, training with a teacher – “do as I do” – is used.
For example, when a robot is trained to diagnose diseases, it is shown a large amount of data on symptoms, tests, and correctly made diagnoses. It treats this knowledge as true and can often surpass its teacher because it learned from a number of examples that none of the real doctors can even see.
It is bad if wrong diagnoses are used for training. The robot sometimes detects them in the learning process and the teacher excludes such examples from learning materials, but often the person believes in his absolute rightness, and the robot inherits his mistakes. In a number of tasks, a person immediately proves to be helpless against the complexity of data he has. Then he builds AI that can help understand the structure of this data, find the data that falls out of general patterns and describe these patterns in some way by means of establishing links, common constructions.
If we talk about building AI based on self-learning, then today such AI relies on the total number of data available to it provided by the person and the person’s understanding of the difference between patterns and chaos. Currently, a self-study is associated with teaching a certain class of models that a person puts forward as the most suitable ones for describing the subject area in which he would like to use AI.
Finally, based on its results, reinforcement learning is an interesting methodology of teaching AI by immersing it in an environment of further use. At the same time, artificial intelligence is encouraged or punished for every action it performs. The simplest illustration here can be training AI various games. Games can be repeated an unlimited number of times, which allows AI to train using very limited data on how the environment is described in which it operates from a person’s point of view.
Recently, one of the main achievements in the field of artificial intelligence is the development of AI, which learned to play Go during the game with itself and began to beat the champions. This approach resulted in the creation of AI for teaching other AI. And finally, the answer to the question: what tasks are most appropriate and effective to use AI for? The fact is that it is absolutely impossible to use AI where no intelligence is needed at the moment.
The key areas of the AI use in the way it is now understood are: managing many coordinated processes (for example, transportation, manufacturing of material objects, energy), helping people quickly to maintain their health and lives (healthcare, rescue works), the legal sphere where independence of decisions from human failings is important.
Today healthcare clearly demonstrates a good adaptation of AI. It is precisely AI that allows not only giving recommendations when diagnosing diseases with any complex method, for example, functional MRI, but also objectifying an approach to diagnostics for the first time based on the simultaneous examination of a large number of indicators and records – a holistic approach. The application of Watson has moved from the research phase into the phase of the national program in South Korea. The first successful project was implemented in September 2016 when artificial intelligence was adopted in the Gil Medical Center. The first South Korean patient to be treated with the help of IBM Watson was Tae-hyun Cho. At the time of examination, the patient was 61 years old. According to the colonoscopy results, he was diagnosed with colorectal cancer. The need for chemotherapy and the selection of the most effective drugs were the reason to introduce data into AI. The specialists and the patient were satisfied, because as a result of the quality treatment prescribed by the machine, it was possible to destroy the cancer cells remaining after resection. In the near future, Japanese doctors will massively use a system of visual diagnostics based on artificial intelligence.
These systems will be able to find such tiny anomalies that doctors would never notice themselves. In Japan, several similar projects are being developed. LPixel, associated with the University of Tokyo, is developing a system that defines aneurysms. Aneurysms can cause brain damage, and its early detection is crucial for the prevention of serious consequences. The company provides its diagnostic AI together with a database of MRI images collected from 10 medical institutions. Doctors are shown a 3D image of the blood circulatory system, on which the algorithm marks the place where the aneurysm can be located with a red marker.
LPixel plans to enter the market with its system in 2019. The research company Fuji Keizai claims that the Japanese medical AI market will grow by more than twice by 2020 – from $ 32.8 million in 2016 to $ 88 million. Behind this growth is the predicted shortage of doctors, the authors of the forecast explain. To reduce healthcare costs, the government will promote early diagnosis and prevention of diseases. An important role should be played by AI. This will lead to the situation when one doctor can work with a large number of patients while maintaining a high quality of work. The fact that a large amount of visual medical data is produced in Japan contributes to this situation.
Recently, an area known as “Fintech” has begun to invest heavily in the development of AI systems. There are already a few successful projects. For example, the use of AI to detect theft. This project is based on the analysis of the AI flow from millions of banking transactions. For a person to notice something suspicious in such a volume of information is simply impossible. And no conservatism can reduce the pace of AI introduction in the financial sector. After all, even the strongest opponents of cars once changed them to horses.
We should not forget about transport. The outstanding success of AI in supporting the fourth-level autopilots is already visible in this sphere. Today unmanned vehicles are a reality. The next step in the introduction of AI in the transport area is the development of transport systems based on the exchange of information between vehicles and objects of the environment: road signs, traffic lights, bridges, structures, road sections. No other technology except for AI can be a quality solution for coordinated and secure maintenance of such systems. The common name for this development direction is Connected Vehicles.