What do you really think about AI for non-researchers?
In recent years, research on AI (artificial intelligence) and machine learning has developed, and it has come to appear in daily news and conversation as well. However, if it is not specialized researcher and engineer, even if it is said to be AI, there are many people who do not know at all about with the recognition of "It seems to be something amazing", "I won the professional with go and shogi". However, it is a waste that you can not understand about AI just because you do not have knowledge. For such people, "What are the points that people other than AI researchers and programmers really think about AI?About Google's AI relationship work on GoogleYonatan ZungerThe column that Mr. explained explainedMediumIt is published in.
Asking the Right Questions About AI - Yonatan Zunger - Medium
Mr. Zunger says "Extreme discussion has been continued over the past few years as to whether AI will save the world or lead the world to ruin." However, this argument seems to be totally nonsense, being the same level as the extreme argument that "SNS collapses democracy" "People can not burn bread due to the spread of toaster". "The discussion that is actually required for the current AI is not a pole theory like" AI is a friend of mankind or an enemy "and it is not a very technical technical challenge because the general people are enough It is a question that can be thought of, and it is also a question to participate in the discussion. "
According to Zunger, we need to understand the point that "what is AI doing" first. AI is a device that "predicts actions to take" based on the information that they have with respect to the input information. Sensors such as cameras and microphones that gather information from the world are distinguished from AI itself, and the device that receives information of "actions that AI predicted" and actually moves to action is not AI itself. If compared to human beings, sensory organs such as eyes and ears are sensors, brains thinking according to external stimuli are AI, devices in which the arms and legs moving according to the stimulus move into action.
As an example of a simple machine that responds to the stimulus it receives, Mr. Zunger lists valves that regulate gas and fluid pressure. The valve receives the pressure applied to the valve body and mechanically determines whether to loosen or tighten the valve. In this case, the valve returns as one output for one input. Zunger says that what AI is trying to do is to input a huge amount of information and determine the output, unlike valves. "If you are an automated driving car you can judge appropriate driving based on information such as the flow of surrounding cars, weather, obstacles, noise, etc. If you are a survey on the ecology of wombat, consult various webpages and summarize the appropriate information It is said. "
byryan and alyssa
Next, it is necessary to know "the advantages and disadvantages of AI". Mr. Zunger points out that one of the advantages AI has is "getting tired and not concentrating". Even if you learn a huge amount of people you will get bored and tired on the way, but AI can input new teaching materials continuously for a long time. Also, you will not be disrupted in concentration during driving of an automated driving car. This can be said to be a very good point, whereas human drivers lose their concentration with a one or two hour driving.
Mr. Zunger lists what "human beings are doing for granted" as a point that AI is not good at (thought to be) a point. If you let AI to do what human beings can do commonly, it may be thought that "AI is not a big deal" by struggling fiercely, but according to Zunger, this is a hurdle in the field that human beings can intuitively do Because I'm raising. This is face recognition that has been talked about recently, it also applies to simple exercise functions that walk and run. In most cases, humans can distinguish family, friends, and TV talents' face from each other without being taught by anyone. However, the face recognition function of AI sometimes erroneously recognizes brothers and parents whose faces are similar, which is not perfect. "Even if you move a robot by AI, it can be said that human exercise functions, such as going through a messy room without difficulty and walking while holding a glass with water, are not good as compared with humans Shu, "he said.
It is said that AI is good at "achieving clear goals in a predictable environment". "It's applied to simple programs and monitoring functions," Zunger said. "It will be slightly more difficult to achieve" a clear goal in an unpredictable environment ", but this is a challenge currently faced by automated driving cars and autonomous drone, AI is the most difficult "A situation where the goal is not settled in an unpredictable environment." In such an unstable situation, AI can not judge what you should do by myself, you can not demonstrate your power, "Zunger He has summarized AI's fields of expertise and weak points. In other words, even if AI gives a vague instruction to "make a goal for life", for example, AI can not make a decision under circumstances where it is not known what goal is to be made the right answer.
It is an ethical question that Mr. Zunger brought out next. One of the famous arguments is "Automatic driving car controlled by AI will crush the car to the shoulder to kill the person in front as it is in order to protect the safety of the passenger when it is about to hit the person in front or to protect the person in frontThere are things like. Mr. Zunger firstly understands in this discussion that "most accidents are caused by accident or negligence of the driver in the first place". Mr. Zunger believes that most accidents can be avoided if the driver is sure of the safety of the surroundings or if he does not concentrate.
However, accidents that can not be avoided absolutely with a small number occur. And the point that Mr. Zunger should discuss before from here is "what program we will apply to AI". "Since AI operates according to a pre-program, we declare" whether to preempt passenger safety and to kill immediate people "or" crash the car to protect people in the front ", a common intention in our society It is necessary to make a decision. " It is not "what AI does" but rather "we should decide how we decide".
And Zunger says he needs to know about "pitfalls of machine learning" essential in the process of acquiring intelligence by AI. "In explaining this pitfall, in a perfect incident"Three teenagers incident"there is. This is a case where a high school student announced that there was a clear deviation in the displayed image when searching Google with the word "three white teenagers" and "three black teenagers." Three white teenagers "showed mainly the photographs of three white people in a healthy and cheerful atmosphere, whereas in" three black teenagers "a face picture taken when arrested (Magshot There was a difference that it was mainly displayed. "
This does not mean Google search is racist. The reason why Google search displayed such an image is that as a result of referring to the text overflowing on the net, there are few white mug shots related to "three white teenagers", "three black teenagers "There are three black men's shots in relation to the cause, it is cause. When people write the text of the news, even if three white teenagers were arrested, only "three teenagers arrested" and only three black teenagers were arrested, "three black teenagers arrested" when three black teenagers were arrested And Google search only followed it. In other words, it is evidence that AI displayed such discriminatory search results is a discriminatory awareness on the human side.
As an ethical problem related to AI, "Gorilla · Tagging Case of Google Photos"Also made a great impact on people all over the world. This was a case where AI recognized the black image as "gorilla", and the developers have developed into a situation where the developer apologizes. Actually, Zunger is the chief of the Google social division who apologized in this case. Even as of January 2018 when more than two years have passed since the incident, GoogleBlock images of gorillas from search resultsAnd it should be a bitter memory for Mr. Zunger, but Mr. Zunger also touches on this case in the column. Regarding the fact that AI judged blacks as gorillas as well, there was no discriminatory feeling in AI itself, so it is said that it is difficult for AI to recognize the face that humans do as they should reason.
Based on these problems, Mr. Zunger says, "After all, what we should understand is" AI is not perfect, it is not a self-conscious person like human being, it's just a tool It is "It is that it is not." Mr. Zunger also told about "the unwarranted drone bombers carrying the AI to kill people" which is often talked about, "Mr. Zunger told the story," With regard to this discussion, AI has never been responsible for making bombing decisions It is apt to fall through the argument that the bombing was decided on the side of the person who is in operation and AI plays an auxiliary role such as assisting Drone's autonomous flight and matching the aim, It is nonsense to take responsibility for the bombing. "
Looking back on the point that "AI can not cope with problems with no clear goal", you can see that what AI does is left to our hands. We need not think about the "threat posed by AI to human beings", but think about "what to do with AI?" "How to operate AI?" This is not a problem imposed only on AI researchers and engineers, but also we have to deal with the influence of the upcoming AI together and discuss it.
in Science, Posted by log1h_ik