Directory of RSS feeds

RSS feeds in the directory: 374

Added today: 0

Added yesterday: 0

Company / Other

Thanks to deep learning, work can be a lot of cars

The scientific approach 31.03.2016 at 09:45

Site news

18-time world champion Lee SEDOL AlphaGo learns something new — the defeat

on 2 December 1942 a team of scientists led by Enrico Fermi observed returning from lunch, as humankind has launched the first ever self-sustaining nuclear reaction within the piles of bricks and boards under a football field at the University of Chicago. Known as "Chicago pile-1" experiment was celebrated in the quiet of a single bottle of "Chianti": those who were present, the significance of this event for humanity, it was clear without words.

Now there was something new, once unnoticed forever changed the world. If the word whispered in a foreign language: you like it, and I heard, but its true meaning has eluded you. Nevertheless, it is vital to understand this new language and the meaning of words that we are told, because the consequences will change everything. Even the most obvious aspects of the functioning of our globalized economy, as well as the ways by which we humans exist within it.

the Language is that a new kind of machine learning known as "deep learning" and "whispered words" — using his computer for, seemingly out of nowhere who took victory over three-time champion th Fan of Cock, and not once, but five times in a row without a single defeat. Many of those who read this news, I decided that it is impressive, but it doesn't even compare with the match against Lee SEDOL, which is widely considered to be one of the best living players, if not the best. Introducing such a great duel of man and machine, the best player of China predicted that Lee won't lose a single game, and he was sure I was going to lose not more than one.

What happened actually? Whether lost four games out of five. And the AI is called AlphaGo now the best player in the world: he was awarded the "divine" grade to ninth Dan. In other words, his level of play godlike. Th officially surrendered the car, just as Jeopardy was subdued Watson and chess — Deep Blue.

So what is th? It's very simple: imagine that go is super-mega-ultra-chess. This achievement may still seem insignificant, another feather in the hat of the machine, because they continue to show their superiority in entertainment games, but this is a considerable achievement, and occurring no longer just a game.

the Historic victory AlphaGo — a clear signal that we have moved from a straight line to a parabola. Progress in technology at present is so clearly exponential in nature, we can expect a lot more overcome milestones than would be expected in any other case. These exemplary achievements, especially in the form of achievements AI specialized in specific tasks that will leave us totally unprepared to them as long as we continue to insist that the work is the main source of income.

All this may sound exaggerated, so let's go back a couple of decades ago and look at what enthusiastically worked in computer technology in the context of human employment:

Source: St. Louis Fed.

Consider this graph as follows. Don't be fooled into believing that talking about the automation of labour we are talking about the future. The future is already here. Computer technology eats our jobs since 1990.

Routine job

All work can be divided into four types: routine and variety, cognitive and manual. Routine work — all the same thing day in and day out, while his solo work tends to change. Within these two types is the work that requires efforts of our brain (cognitive) and work that requires efforts of our body (hand). Where all four type once started to grow, routine work is stuck in one place in 1990. This happened due to the fact that routine work is the most simple to perform technological devices. For work that does not change, you can write an algorithm, and a machine with such a job cheers.

it's Sad that this type of work — routine — once formed the basis of the American middle class. It is routine manual work changed, Henry Ford, started paying people for it the salary of the middle class, and it is routine cognitive work once invaded American offices. Jobs of this kind are now becoming more and more inaccessible, leaving only the two work rainbow future: the work that do not require a lot of thinking — we pay little; and jobs that require such mental stress that we pay people well so they thought.

If you imagine our economy in the form of an airplane with four engines but can fly on two, while they work, we can avoid the drop. But what happens if you refuse and also the remaining two engines? This threatens the success of the advanced areas of robotics and AI for the two remaining engines, because for the first time in history, we successfully teach machines to learn.

neural networks

At heart I'm a writer, but my education is in psychology and physics. I capture both these Sciences, so as a student I was focused on the physics of the human brain, in other words — cognitive neuroscience. I think once you start to study how human brain works, as masses of interconnected neurons somehow form what we describe by the word "mind", everything changes. At least, everything changed for me.

Here's a visual aid of how our brain works: it is a huge network of interconnected cells. Some of these connections are short, some long. Some cells are only connected to immediate neighbors, and some connected with many. Electrical impulses pass through these connections at different speeds, in turn triggering the following cascade of pulses. Like falling dominoes, but only faster, bigger and much more complicated. Surprisingly, the result is we, and what we learned about how we work, now applied to the machine.

One of the results of such fitting is the creation of deep neural networks — a kind of reduced version of the virtual brain. They provide the possibility of machine learning, which has made incredible advances that were previously considered out of reach much more, if at all possible. How? This is not just the obvious ability of our computers to the growth and expansion of our knowledge in the field of neurobiology, but also the rapidly growing volumes of collective information, also known as big data.

Big data

It's not just a bunch of pretentious words. This information, and when it comes to information, it becomes clear that every day we create more and more. In fact, we create such volumes that the SINTEF report 2013 States that 90 percent of all information in the world was created in the last two years. This incredible figure is doubling every one and a half years: thanks to the Internet, where in 2015 every moment we "like" 4.2 million photos on Facebook, 300 hours of video uploaded on YouTube and sent 350 thousand tweets. Everything we do generates information as never before, and more is exactly what you need machines to learn. Why?

Imagine that the computer is programmed to recognize the chair. You need to enter a ton of instructions and as a result this will be a program that detects the chairs that are not chairs at all, and the chairs she did not detect. So how do we ourselves learn to recognize the chairs? Our parents pointed to the chair and said, "That chair". So we thought that this whole problem is solved with chairs, so we poked in the finger table and said, "That chair", and then the parents said it was the table. This is called reinforcement learning. The label "chair" becomes linked with every chair that we see, so that some neural pathways are impacted and others are not. To start in our brain response "the chair" should be close to the sites of our past experience. By and large, our life experiences — the vast array of data filtered by the brain.

Deep learning

the Power of deep learning is that it is a way of using a large amount of information to teach machines work without clear instructions, that is the way we do. Instead to describe the "essence stolow" computer, we just connect it to the Internet and show him the millions of images of chairs. Then he will be able to get an idea about this "entity". Then we will test him even more images. When he makes a mistake, we fix it, that will only improve his skill of identifying chairs. Repetition of this process leads to the fact that the computer knows what a chair is when he sees it, as good as we can.

An important difference is that, unlike us, he can view millions of images in seconds.

This combination of deep learning and big data has led to incredible advances just in the last year. In addition to the amazing results obtained AlphaGo, artificial intelligence, Google's DeepMind learned to read and understand what he reads, after studying hundreds of thousands of news articles. DeepMind also independently learned to play dozens of video games on the Atari 2600 is better than people just checking the score on the screen and playing over and over again. AI called Giraffe in a similar way taught himself to play chess, using a dataset of 175 million chess positions, reaching the level of International master for 72 hours, continuously playing with yourself. In 2015, the AI was even able to pass a visual Turing test, studying so that he could identify the unknown letter in an imaginary alphabet, and then to immediately reproduce the letter so that it was completely indistinguishable from what is written in the same job man. These are essential milestones in the history of AI.

However, despite all these milestones, when the question was asked how soon a computer will be able to win the outstanding player in th, a few months before Google ads about winning AlphaGo experts was given the following answer: "Perhaps in the next 10 years." The decade seemed like a good option, because th — the game is so complex that I will give the opportunity to describe it to Ken Jennings from the Jeopardy game, another former champion, slain AI:

"it is Well known that go is more complex than chess, with its huge blackboard, longer games and a large number of figures. The team of programmers AI DeepMind Google likes to say that in all the known Universe of game options go more than atoms, but this greatly underestimates the problem of computing. Positions in the game of go is 10 to 170 degrees, while the atoms in the Universe is only 10 to the 80. This means that if there are as many parallel Universes as there are atoms in our (!), the total number of atoms in all those Universes combined would come near approached to the number of possible positions on a go Board"

This bewildering complexity makes it impossible to brute force every possible position and every possible move to determine the best. But deep neural networks bypass this barrier the same way as your own brain: learning to evaluate the course and choose the best. We do this through observation and practice, and similarly it makes AlphaGo: analyzing millions professional games and playing with himself millions of times. So the answer to the question when will fall in front of the car was not even close to 10 years. The correct answer is "anytime".

At any time. That's the new answer 21st century to any question associated with when the machine can learn to do anything better than humans, and we should try to listen to it.

We should recognize what it means for exponential changes in technology first-ever access to the labour market varied. Machines that can learn anything — this means that any person in danger. From hamburgers to health care, the machine can successfully perform these tasks without assistance or with less assistance of man, at a lower price than people.

Amelia — just one of many AI, which right now is in beta testing. This AI was created by the company IPsoft for the past 16 years, she learned to perform the work of employees of the call center. It can seconds to learn what takes us months, and she can do the job in 20 languages. And since she is able to learn, over time, it can even more. In one company, where it had been tested, she has successfully treated every tenth challenge for the first week, and by the end of the second month it was decided that six of the ten questions. And because of that she can leave without work 250 million people worldwide.

Viv — AI from the creators of Siri — will be released very soon, and she would be our personal assistant. It would take us to task online, it will be like a Facebook news feed on steroids: will offer a to consume media, which in her opinion we should be like. In addition to this we will see far less advertising, and this means that the entire advertising industry — the industry on which the entire Internet would significantly falter.

the World with Amelia and Viv — and countless more AI, which will soon come online in combination with robots, such as Atlas is the creation of a new generation of Boston Dynamics, is a world where machines can do the work of any of the four types, this means that there will be serious social changes. If the machine can do the work instead of a person, whether the person to risk their material well-being for her? Should the income be tied down to the business of employment, and the availability of work to be the only way to generate income, while jobs in many completely not available? If machines fill up more of our jobs, and they don't pay for it, the money where to go? And how will their purchasing power? Is it possible that a lot of our work that we have created, should not exist, but exist in order to get the money? We must begin to ask these questions, and fast.

Separating income and work

fortunately, people are starting to ask these questions, and the answer there is that built in his eyes. The idea is that the machines worked for us, but we worked those, whose work we believe is really important, just a person on a monthly basis providing monthly income regardless of work. This income will be guaranteed for every citizen no matter what, and it's called — a universal unconditional income. Taking it on Board, in addition to protecting from negative consequences of automation, we will also reduce the risks inherent to entrepreneurship, and reduce the extent of bureaucracy necessary to increase revenue. For these reasons, it is supported by people with diverse political views and is already at the stage of possible implementation in countries such as Switzerland, Finland, the Netherlands and Canada.

the Future — a world of accelerating change. It seems unreasonable to continue to see future similar to the past, where new jobs will be simply because they have always been before. WEF beginning of 2016, opening 2 million jobs and eliminating 7 million. This loss of 5 million jobs. In the often quoted work, published in Oxford, predicted full automation half of current jobs by 2033. At the same time the cars are on autopilot, again, thanks to machine learning, can dramatically affect all economies — especially the economies of the USA, as I wrote last year about the trucks on autopilot, is eliminating millions of jobs in a short period of time.

And now even the White house in a stunning report to Congress stated that with a probability 83% employee, who in 2010 earned less than $20 an hour, will give my work car. Even people earning up to $40 an hour, you have a chance of 31%. To ignore the probability of such — that is, to use a ridiculous strategy of "duck and hide" to prevent the effects of nuclear explosions during the Cold war.

All of this tells us why now is exactly the most knowledgeable in the field of AI enthusiastically sounding the alarm and talking about absolute income. During the discussion at the end of 2015, at Singularity University, an outstanding data management specialist Jeremy Howard asked: "do You want half of the people on the planet starved simply because they literally cannot bring economic value or not?" before you assume: "If your answer is "no", then the most reasonable way of distributing wealth will be the introduction of a universal unconditional income."

the pioneer in the field of AI Chris Eliasmith, Director of the Center for theoretical neuroscience, warned about actual problems of influence of AI on society in an interview with Futurism: "AI is already affecting our economy, I suspect more countries will follow the example of Finland in the introduction of an unconditional guaranteed income for people."

Moshe Vardi expressed the same idea in 2016 after speaking at the annual meeting of the American Association for the advancement of science about the emergence of smart machines: "We need to rethink the very basis of our economic system... we may need to consider the introduction of an unconditional guaranteed income."

Even the chief scientist of Baidu and founder of project deep learning "Google Brain" Andrew ng in an interview from the scene at the deep learning Summit this year expressed a common view that an unconditional income must be "seriously considered" by governments, citing the "high likelihood that the AI will create massive movement of labour".

When the creators of the tools begin to think about the consequences of their use, should wish to use these tools to listen with great attention, especially when at stake is the lives of millions of people? If not, then what about the Nobel laureate economist who agrees with proponents of guaranteed income?

None of the countries are not yet ready for the changes waiting for us. Unhelpful behaviour-paid labour leads to social instability, and the lack of consumers in the consumer economy leads to economic instability. So let's ask ourselves: what is the purpose of the technologies that we create? What is the purpose of the machine, which can operate independently, or AI, which can perform 60% of all our work? Let us work more for less pay? Or allow ourselves to choose how we work, and to refuse any salary, which we consider insufficient, because we already get the money devoid of cars?

And what important lesson can be extracted in an age when machines can learn themselves?

I'm just saying that work for the machines; for a human life.

Author: Scott Santens

the Original: Medium