Why do companies want to collect data on user behavior can be understood from the mechanism of bots monitoring humans on the net

It is said that user behavior on the Internet is a treasure trove of precious data for creating a mechanism to maximize profit. A movie that explains through a machine learning what kind of mechanism a bot existing in the Internet functions in the Internet "How Machines LearnYou can understand why companies are getting more fever with data collection by looking at "the structure that trains computer algorithms and maximizes profit from human behavior".

How Machines Learn - YouTube


The world of the Internet is full of "algorithms".


Perhaps because I am looking at this movie, the algorithm recommended.


If you click on the movie, the algorithm will note it.


If you open Twitter, the tweets you should see on the timeline are displayed ......


When there is a desired image ......


The algorithm tells you the perfect image.


The price of on-line mail order is also determined by the algorithm ......


Algorithm follows for remittance procedure ......


Algorithmic trading is rampant on the stock market.


But you do not know how the algorithm works.


Initially, the algorithm was made by humans.


I instructed the computer in an understandable way to do "some behavior" when "something" happened.


However, the big problem is that the conditions imposed on the algorithm are too complex to understand any more.


An enormous number of financial settlements are done in one second ... ...


Among the huge number of YouTube movies, things that are worth seeing are sorted out.


Aircraft tickets ......


The highest price a user pays is presented by the algorithm.


The algorithm decisions are not necessarily correct.


But it is far more accurate than the decision that humans make.


How this kind of bot algorithm works is no longer understandable to people who made the algorithm.


Furthermore, companies do not reveal detailed contents of algorithms.


Because the algorithm that produces a large profit is "high salary".


The contents of the algorithm are black boxes.


Furthermore, the science of state-of-the-art algorithms is also difficult to understand in the first place.


So, here we are not thinking about how the brain of the algorithm works, but we will consider how the results are brought about, ie the mechanism of "machine learning".


Here is a picture of a bee.


Another thing, apart from bees, there is a photograph of the number "3". Consider a scene that distinguishes between 3 botches of bees and numbers as a bot.


These distinctions are very easy for human beings.


Even small children can easily distinguish between bee and number 3.


However, it is extremely difficult to teach these distinctions in Bot language so that they can be understood by Bot.


Therefore, rather than making bots that can be distinguished ... ...


I will make "a bot that can make a bot".


It is efficient to leave the production of bots to "bots that can make bots".


Furthermore, if you make teacher Bot teaching bot ......


You can educate bots efficiently.


The existence of such "bot that can make bot" and "teacher teaching bot" like this makes the work that a human programmer should do is simple.


At first, the brain 's intracerebral circuit is randomly created.


And the suitably produced bot will be sent to the teacher bot.


Nonetheless, it is still difficult to create a teacher bot so that bots can be educated so that they can distinguish between bees and numbers 3.


Therefore, humans decide not to teach to teacher Bot but to give a large amount of images. Images of bees ......


Give the teacher Bot a large number of images of the number 3.


And we will give key things to distinguish between bee and number 3.


The important thing here is that even a teacher Bot can not educate a bot.


Teacher Bot will "test" instead of educating.


Are you stupid bots ......


I will get bad grades in the test.


Some bots will have very bad results.


This is not a bad bad. It is only made to be so.


From the test results, distinguish between bots with high scores and bots that are not ... ...


Bots with low points went to the trash box. It will be rebuilt again.


"Bots that can make bots" can not make bots well as ever, but you can duplicate the remaining bots and combine some circuits to remodel it.


And the bots will be sent back to the teacher Bot again.


Teacher Bot tests again ... ...


The bots that were screened by the test are duplicated and remodeled.


It was tested ......


The bad bot is recyclable. Repeat this work.


In this iteration, excellent bots will be elected.


And, except for the most excellent bot, I went to the trash box.


In old-fashioned schools, the number of students guided by teachers is limited.


However, in the world of machine learning it is OK at all that the number of students is thousands.


The test is not necessarily 10 questions ... ...


You can have millions of questions solved.


In the world of machine learning, an enormous number of bots are selected ... ...


The test is repeated. There is no limit to the number of tests, and it is repeated as many times as necessary.


The bot that first survived was merely lucky.


That lucky bot is remodeled ......


Luckily well remodeled "improved" within the iteration that bots are chosen ... ...


An enormously lucky and excellent bot survives.


After the selection by the test and remodeling in this way, the probability that the bot can distinguish bee from the number 3 is gradually increased.


Bots can now distinguish images with a high probability of 77%.


Of course it is not enough yet.


A bot born from a place that is an infinite warehouse and a slaughterhouse ... ...


You can distinguish surprisingly well the images of bee and number 3, which you have not seen before.


However, why Bot can successfully distinguish images in this way is unknown to Bot himself, as well as Teacher Bot, Bot that can make a bot, even human beings.


It can be said that the link on the brain of the bot born by repeated improvements has a level complexity which can no longer be understood. Even though you can elucidate the code of individual links, the intertwined complex collection of links go beyond human understanding.


However, although it is an excellent bot, it is only the subjects tested that can be distinguished. It is impossible to distinguish bees from bees.


Upside-down images ......


We can not distinguish misleading images well.


As usual the teacher Bot can not educate the bot ......


What humans can do is just to add a problem to be tested. It is ok if you add a problem that even the best bot will be mistaken.


The above indicates the reason why companies want to track and investigate user behavior.


"More data" is a "longer problem", that is, it creates "a better bot".


Image selection work used for "proof of not robot" used in CAPTCHA etc ......


Actually, there are aspects that make human beings act to train bots so that they can distinguish between horses and humans.


Recent CAPTCHA should have many problems to distinguish cars and road signs. This is to train the image recognition ability of the automatic driving car.


So, what is going on to make the YouTube movie viewable to the user for as long as possible?


It is easy to measure the time that users stay on the site to watch a movie.


Let each bot stick to the user and record what kind of movie you watched each time.


The result of the longest movie comes out.


After the test results come out, the role of bots and teacher bots that can make bots.


Just as test and remodeling are repeated.


The bot that survived in this way ......


You will be able to offer recommended movies that can attract users to movies.


Behind the recommendation function of YouTube, there are countless bot choices being made.


There is no way to know what is going on there.


The only thing is sure that the new bot has higher precision than the previous bot.


There are many mechanisms in the world of the Internet that make users react more.


And "points" that profit is the highest are being discovered.


The human being who produced the bot ......


It is not a mistake to say that the state being manipulated by the bot. Even those who created bots no longer understand the mechanism.


Bot is constantly looking behaviors such as "Like", comments, shares, tweets. In order to create a condition that humans feel more comfortable, it is human beings who wanted the state of overflowing bots.