Smart car features

On our channel, we launched a new column “Future Hunters” ( here is a video ). We also decided to experimentally release the decoding of this video, adapting it as much as possible for the reading format. Welcome under cat.

Now it may seem ridiculous, but this is how a new technological revolution begins for our civilization - machines have learned to learn. We present you a pilot release of the new Sci-One format “Future Hunters” - together with scientists and engineers who create technologies that are ready to change the world, we understand how this happens, learn what works and what does not work and what is the problem and sometimes we reflect on how we and the world are changing with the development of technology. Today we are introduced to machine learning. And help us in this Andrew Sebrant .
')
Andrei Sebrant: “ The story about learning is the essence of machine intelligence. We have learned to create such programs and run them on such hardware that this hardware-software complex begins to have the ability to learn independently . ”


The drone (at gif above) survived 11,500 collisions in order to never repeat its mistakes.
No external management - he suffered. The learning algorithm in general is simple -
you start and fly forward. If you collide, you fly back to the starting point and repeat everything again.
But it would have been a meaningless cycle of drone bullying if the neural network had not processed the footage from the video camera on the UAV. Before getting into this network, the frames were divided into two groups, some describing what was the environment before the collision, others - during. Then the neural network understood when it is good and you can fly, and when it is bad and there will be a blow. This is how the drone flew 40 hours into 20 different rooms and now nothing hurts anyone, of course, only in a familiar environment.

Andrei Sebrant: “ On the one hand, we also call a chess game an intellectual game, but since the computer beat Kasparov, no one was particularly excited about artificial intelligence, but around Guo they are now excited.

The creators of the algorithms say, “We never could have calculated the batch to the end in this place, we had to lay something non-algorithmic in the machine so that it would learn by itself”. And it is precisely in words “she learned herself” that people start to go crazy because it seems that the ability to learn , and not just the ability to perform a very complex, very long, pre-formed by man, sequence of actions, is what she (the ability to learn) distinguishes man from car . "

The historic match took place between one of the world's best gay players, Lee Sedol, and the AlphaGo artificial intelligence system. She won 4 games and lost 1, and this meant that for the first time the car was better than man in this ancient and very difficult game. Many experts believed that this would happen no earlier than in 10 years.

Two moves that Lee Cedol and Alpha have changed the future
In the game number 2, the Google car made a beautiful move that no one would make. In the eyes of the whole world, this move just perfectly demonstrated the incredibly powerful and even incomprehensible talents of modern artificial intelligence.

However, in the party number 4, a person has already made a move that no car could expect, and is also very beautiful. In fact, he was as magnificent as the car’s move in batch 2, no more and no less. And this move showed that despite the fact that now machines are capable of moments of genius, it is unlikely that people lost the ability to find their own insights. And it seems that the development of such machines will in the near future lead to the development of a human genius in tandem with our creations.

Wired magazine, March 16, 2016

Andrei Sebrant: “ The simplest way to explain the principle is to describe the old-old problem" How to distinguish a cat from a dog? ", Because in fact in the traditional paradigm of computing technology, you must first understand with the power of our mind that in a photo of a cat makes it in our eyes cat Here you can begin to describe an object of class 'eyes' and it will be a long description “this is some kind of object with glitter, most often round, in the middle it contains dark space, sometimes round, sometimes extended. And this object is the 'eye' and if this space is round, then, most likely, the eye is doggy. And if this dark space is elongated, then, most likely, the eye, feline. " And this is one of the many signs that we will now formulate and by which one can distinguish a photo of a cat from a photo of a dog. But it is clear that from the fact that a cat or a dog has closed its eyes, it does not cease to be a cat or a dog, and therefore we will have to continue to describe “whiskers stick out differently” and describe the concept of what is “mustache”, as the algorithm can distinguish “mustache 'in the picture and continue to understand this feline mustache or dog, but then, nevertheless, even with a shaved mustache the cat does not cease to be a cat, and therefore we must consider another class, for example, the shape and location of the ears, describe the class of the object' ear ' and so on.

This is an old classical approach, when we try to understand inside ourselves what distinguishes a cat from a dog in our eyes, then we try to translate it into a description of certain procedures and continue to allow powerful computing technology to analyze the image to highlight these objects and to determine the probability for all parameters ( "eye", "nose", "ear", etc., etc.) is a cat or a dog.

But, in fact, we never did this in childhood as a child, nobody in childhood taught us to do a comparative analysis of eyes and mustaches. We just saw quite a lot of kittens and puppies, adult dogs and cats, and then somehow we quickly learned to distinguish them, not very reflexing exactly how we do it. Now it has become possible to do the same with machines and this is the key moment of machine learning and this, in essence, is the key moment in the emergence of 'machine intelligence', when the machine, which was not explained what cats and dogs are and how they differ, just gave a million pictures of cats and a million pictures of dogs and just set the task to determine the cat it or the dog on the two million first picture . ”

Smart cars seek recognition even in art. In 2016, the first time they fought for the title of the best artist. At the competition Robot Art, the main prize of $ 40,000 was awarded to PIX18 - a decommissioned industrial robot hand endowed with a neural network.
If another smart machine is given to read a lot of poems, then it can write its own.

Andrei Sebrant: “ “ Why did she decide that? ”- but what the hell knows! This is also what scares people the most. “How did we teach her if we didn’t explain what exactly should be distinguished?” - just like our children. We did not explain to them how to distinguish a cat from a dog, we just showed them cats and dogs. And this story has led to the emergence of this special phenomenon “trained machines”, machines possessing 'machine intelligence' because above I, in fact, described a very popular task - the task of classification. So let's classify objects of the cat or dog class into these two groups. What happens if you show your portrait to the same algorithm? - he has no other choice, he is not trained in the fact that there are some other objects in the world besides a cat or a dog. Further, it will be over, it will be interesting what he will consider you. And then you probably would be wondering why he took you for a cat, and not for a dog. But the program cannot accept another decision because it does not know how! In general, in her knowledge there is nothing but cats and dogs. It is clear that further begins the interesting task of teaching the diversity of the whole world. That is, in the end, of course, good, modern programs distinguish between hundreds of thousands of objects, among which anything - steamships, houses, people, umbrellas, clouds and anything. It took them quite a lot of time, but this time they had and need to understand that the computer also learns much faster than you and I.

Therefore, it is at that moment when it suddenly turns out that the computer begins to distinguish, identify objects in the picture, and then (if the computer also read the signature signatures) says that “this picture shows not just the ocean” and is not just a lonely figure in the rays of sunset, but “this picture depicts loneliness” (yes, there are programs that can write this when you show them your favorite photo of a lonely girl at sunset) - at this moment people have something like “ so, it is generally abs An interpretative concept, where did he get it from? ”- yes, from our experience, because there are a lot of pictures on the Internet with approximately the same set of objects, next to which the text about loneliness is not enough.

That is, it turns out that the machine can learn from all the wealth of information that is now in the network, and there, as you understand, there is everything, starting with Wikipedia, ending with our photos with you, open posts in social networks, etc. etc. ".

Your search queries can also be taught by cars:

How to learn to draw
Blew neck ointment
Guf Christmas download
Synonym for the word connection
Dzerzhinsk aspening route
How to add to private
New Year's Eve Ideas
Watch let them say

Yandex. Autopoet

This poem was written not by the neural network, but by a regular program, but if you “listen to it performed by the author,” you will hear how the machine read it, which itself has learned to speak. To this end, she listened to 300 hours of male, female, childish speech and synthesized her own using neural network.

Andrei Sebrant: “ It turns out that, as a matter of fact, (this is not my phrase, but the phrase of many people who stand at the basis of the machine learning algorithms themselves)“ machines begin to know the world in which we live with you ”. Moreover, to learn and learn in it at least for the time being to classify, i.e. distinguish any objects from each other. This is machine learning. How much is intelligence? - Well, let's see what will happen next, because such a great knowledge and understanding that this picture shows a sunset, which people associate with loneliness, is not enough for a normal full-fledged intellect which needs a lot more, i.e. we need some personal experiences.

I did not accidentally speak so haphazardly, intricately, in fact, I showed you some kind of mosaic, with pieces. I pursued a definite goal - to show that there is no very clear understanding of what “machine intelligence” is today, including because there is no clear understanding of what intelligence is at all.

These words are used to describe a variety of pieces of a huge, rapidly developing field called machine learning. That's about its individual pieces, a separate application, about what it changes (and it drastically changes not only our life, but also industrial processes, and this is what is called the fourth industrial revolution ) so this is worth discussing separately . "

---

We remind you that this was the decryption of the first video of our new rubric “Future Hunters” (we attach the video just in case):



Write to us in the comments if you like this format (video decoding) and whether it is worth it to continue working on it.

May science be with you.