Developing a self-taught robot: DARPA runs the risk of crossing the line

Defense Science Agency DARPA is preparing to launch an almost 4-year-old project to develop artificial intelligence that can self-educate and improve itself.

The new project is not about modeling the human brain and running the model on a computer. In the short term, this path is considered by many scientists to be a dead end due to the exceptional complexity of the human brain and the little known principles of how our mind works. DARPA plans to go a different way: build machines that will learn using probabilistic programming algorithms, that is, sift through huge databases and choose the best solutions to the problem. In the course of this work, artificial intelligence will learn and after some time will be able to easily solve simple tasks. It is with such projects in fiction novels begins the uprising of cars and the death of humanity.

DARPA experts suggest that commercial campaigns for 46 months develop new probabilistic programming technologies for self-learning machines, abbreviated PPAML. PPAML technology will help the military to more effectively solve many analytical tasks that today require enormous human resources. In particular, self-learning machines will be extremely useful in conducting intelligence, observation, speech recognition, driving a car, sifting information in search of valuable data, etc.

The hardware of artificial intelligence can be varied: supercomputers based on multi-core processors, a network of ordinary PCs and cloud networks.

It is difficult to say how realistic it is to create such a complex artificial intelligence system in the next few years. Currently, there are many automatic analytical tools that facilitate the work of people, but so far none of them are capable of real self-learning. For the time being, the highest achievement of the technology of working with the database essentially remains the comparison of archived pictures with the image coming from the video camera. Therefore, DARPA emphasizes that PPAML algorithms will be much more difficult, in particular, they will be able to combine different types of information, independently formulate a request and formulate a solution based on the data obtained.

Comments

[esaul]

"Prophecies" by H.G. Wells are being realized. It remains only to wish that the artificial intelligence will have "implanted" point number 1 - Do not harm human life and health ....

[Alexander Romanov]

Quote: esaul

point number 1 - Do not harm life and human health ....

If the brain is self-learning, it will quickly cancel this level itself. And with the growth of brains, the computer will understand that with gays it is not on its way.

[KoRSaR1]

If you remember, all artificial intelligence captured the world in films)) In the movie "I am a robot" AI explained its behavior simply, AI did not break the first law on "No harm to health and life" AI evolved and began to take over the world, arguing that we inflict irreparable damage to ourselves, by drug wars, etc. So there is a double-edged sword, and both are not beneficial to us.

[vilenich]

Quote: Alexander Romanov

If the brain is self-learning, it will quickly cancel this level itself.

We need to look at the Terminator again, outline an action plan!
They play out with artificial intelligence ...

[Iraclius]

There is good reason to believe that mental activity is of a quantum nature. And this means that building AI systems on an existing hardware base is impossible in principle. For a decade, quantum computers have not left the walls of laboratories, and all that has been done is machines on several qubits, which is more than modest. A number of scientists generally believe that such a machine is impossible to build in principle.
What kind of AI can we talk about if the mechanism of the human brain is completely unclear?
Systems that mimic human work in some areas, for example, expert systems and pattern recognition systems, actually have nothing to do with thinking and are subject to formalized algorithms.
So we can sleep peacefully - in the next twenty years there are no prerequisites for creating artificial intelligence. Although it is possible that this is bad, because the tangle of contradictions and problems in the world is growing every year and the people themselves are not able to resolve it, alas.

[gridasov]

Alas! You can build a radial numbering system for nine and another count. numbers that do not change the properties of their relationship. Such a system has the property of positioning codes relative to the coordinate system. In other words, the capacity of the system can change without changing the "space dimensions". Remember how many combinations there are in a Rubik's cube. For our system there are no restrictions in numerical terms at all, given that the position of each code is strictly defined, does not repeat and is not subject to ambiguous interpretation. The system has the property of fractal change based only on the summation or subtraction of "zeros". The property of such a system is the lack of proportionality of energy costs for the creation and promotion of code, as in a binary coding system. Since the spatial structure and the codes are positioned, the information capacity changes without dependence on the movement of the codes, which means that the speed, as such, is generally leveled as a concept and property.
Moreover, a transistor concept with a pulse corresponding to each number with a common base has been developed for such a system.
Yes! people do not spend different amounts of time remembering different distances. The system in which the inputs and outputs are located at each of its points has the same property.
The intelligence of a person, like any rational creature, is distinguished by such a quality as the fact that any analysis is done with respect to a variable reference point. That is, any analysis is associated with a certain relative reference point. The same thing happens with the system.
And it doesn't take decades to implement. Everything is much simpler. And such a system cannot be "opened". One can approach these landmarks, but it is not possible to reveal the semantic content, because there is no single vector of the path of analysis.

[Nevsky]

Terminator in the USSR, based on the series "The Chronicles of Sarah Connor" + the series "Guest from the Future"

[Alikovo]

in 5 years the situation will be like in the movie "On the Hook" where artificial intelligence is trying to seize power in America.

[rauffg]

Does anyone know where the "Off" button is at the terminators? It just might come in handy. )))

[Alexei]

Quote: rauffg

Does anyone know where the "Off" button is at the terminators? It just might come in handy. )))

Her. Off button - these are the scientists who are developing this AI. We'll have to move from the future at this very time and bring them all down.

[Egoza]

And then in America there will be a riot of cars .....
But for every "theirs mind" in Russia there is a sledgehammer!

[ia-ai00]

Sledgehammer in the image of a HACKER ...

[Blazard]

Lol I can imagine what capacities are needed for at least more or less normal functioning of these brains and data centers for storing such a database. With the current capacities, the work of FIC is impossible. That's when the first stable quantum supercomputer is built, then you should worry already ...

[Delink]

Blazard, data centers are already there, the Internet. Intel will be able to connect to them without any problems. And filter the Old. I wonder how many languages ​​he will know? And how many talk? And which servers will look at?

[Avenger711]

What info? Zheshechki? All info about the Internet is now at the level of Google search databases.

[Blazard]

And the speed of data exchange? Even if we take as a basis that the entire exchange will take place over multimode fiber with the maximum possible bandwidth - all the same, the losses in data processing time will be much more significant than if everything took place within one complex "DB - Supercomputer". So for now, AI will be limited by the bandwidth of the network and exchange channels.

[vilenich]

Quote: Blazard

I can imagine what capacities are needed for at least more or less normal functioning of these brains and data centers for storing such a database.

And I remember the book "I Want to Know Everything" published in 1964, there was such a comparison that the energy of Niagara Falls would not be enough for the operation of an electronic computer, which in performance could reach the level of a human brain. It is not surprising at that time computers were probably still on the lamps!
Let's wait, progress does not stop.

[Boot under the carpet]

Hmm, I wonder if the principle of the neural network will be introduced there :) Well, well, skynet and the uprising of cars is coming! :-)

[Avenger711]

A neural network is such a dupe.

[Ragnarek]

they’ll make a terminator for themselves! the Japanese strenuously bring Godzilla near Fukushima, s terminators and zombies

[Dwarfik]

Kind! Well, we were surprised at films about artificial intelligence, the uprising of machines, but everything is possible, and who knows, I will no longer be surprised somewhere in a bar when a naked man demands: "I need your clothes!"

[virm]

The article is slightly thickened, of course. No skyneta as a result does not arise. Probabilistic programming, or as it is also called stochastic, is actually not related to artificial intelligence.
This is just one of the methods for finding the optimal solution to the problem. Simply put, if the created system is designed to control a car, then it can to study just that. And that’s all. Maybe, over the course of operation, moving along given routes, this system learn to get from point A to point B. Without accidents. For an acceptable time. But she never learned anything else. Since, just not know about the existence of a world outside the freeway.

[Iraclius]

The first competent comment. The classical problem to be solved in the framework of the theory of combinatorial optimization is the traveling salesman problem, which, when the number of nodes exceeds 66, becomes transcomputing and crude methods of simple enumeration become unacceptable. Among other things, it is solved with the help of the genetic algorithm and the elastic network method, directly related to the mechanisms of establishing ordered neural connections. It doesn’t smell of any intellect, because the usual algorithmization is used.

[Lopatov]

"Combinatorial optimization" is it "game theory"?

[Iraclius]

No, this is a section of optimization theory. Some textbooks claim to be a subsection of multi-criteria optimization. Game theory specifically seeks out optimal strategies in games.

[Lopatov]

Quote: Iraclius

Game theory specifically seeks out optimal strategies in games.

??? As far as I remember, we studied game theory as an application to our specialization. Optimum supply of ammunition. Or am I confusing something, or you.
Although, I agree, I’m a ram in this, we taught VMs with an emphasis on Probability Theory — specificity

[Iraclius]

I myself am not an artilleryman. If you state the conditions of the problem, I can tell by what methods it can be solved. If in this problem the search for the optimal way of transportation was investigated, then I dare to assure that this has nothing to do with game theory. This is the classical Monge-Kantorovich problem (transport problem), which is solved by linear programming methods.
In general, what happens in applied mathematics under the game is understood as the competition of two or more parties for the realization of their interests.
It is possible that you used the application of Markov processes.

[Lopatov]

Ches the word, I do not feel so "savvy" in the VM to argue with you. I will take your opinion on faith. Thanks.

[Iraclius]

Yes, not at all.
About neural networks. In the well-known program Statistica, starting, if I'm not mistaken, from version 7, it is possible to train neural networks yourself - a rather interesting task. There are a certain number of good textbooks, for example "Neural networks. Statistica Neural Networks. Methodology and technologies of modern data analysis" - V.P. Borovikov. (ed.) A good book for general development and all those who are interested in mathematics. Even my cadets did laboratory work on it and were very pleased with the process.

[Averias]

Hmm, that's all doubtful. The Canadians are over there, they say that they have invented a quantum computer a long time ago and even sell it. Only here on any question from the category - how and what - do big eyes, shhhh, they say a secret. That is, they do not give any information about this "miracle" of engineering. And under a terrible secret they keep a list of those companies that allegedly acquired this "super" computer from them. So it is here. Back in Soviet times, we had a working prototype - an optical computer with phenomenal computing power. The West has spent huge amounts of money (on bribes and bribery) to bury this project. The same fate befell the YDMT (low-thrust nuclear engine). And the optical computer, the Japanese tried to collect and restore the documentation bit by bit, did not work. And they were the first to say that the Russians made a revolution in this direction. Wait and see.

[Iraclius]

That’s all nonsense. The company was not able to explain to the scientific community what their computer was like there. This is just such a way of public relations and attracting investment.
For several decades now, IBM has successfully beat Congress funding for the development of cognitive processors with a hardware neural network. Sense - 0 whole, hell tenths. But a steady flow of money for conducting purely commercial research.
In fairness, I note that all serious programs in the field of AI are strictly secret - you can take a word for it - and no worse than nuclear programs in the middle of the 20th century. So ordinary mortals will be the last to know about the invention of AI.

[Iraclius]

Here, by the way, about these guys:

Canadian company D-Wave Systems, engaged in the development and creation of quantum computers, received an investment of $ 30 million dollars.

The group of investors included Amazon founder Jeffrey Bezos and American company In-Q-Tel, which invests in promising information technology projects and supplies the latest CIA and NSA equipment.

“Our intelligence clients have problems that are difficult to solve on a computer with a classic architecture,” comments Robert Ames, one of the In-Q-Tel executives. “We believe that quantum computing can come in handy here.”

It is worth recalling that

The “first commercial quantum computer” D-Wave One, made by Canadians, was sold last spring for 10 million dollars. The buyer was the weapons giant Lockheed Martin, who installed the purchased machine - a black cubic block with a cryogenic system, its size resembling the first computers - at the University of Southern California.

The American manufacturer of weapons and aerospace technology has decided to use D-Wave One in a research project aimed at creating an effective automatic search system for errors in specialized software (errors of the kind that, say, delay the trials of the new F-35 fighter-bomber).

Vesuvius processor from 512 qubits

The D-Wave One processor contains 128 superconducting qubits with Josephson contacts — compounds of superconductors separated by a thin dielectric layer — of the Nb / AlOx / Nb type. In essence, the qubit array is an artificial system of spins from the Ising model with programmable spin-spin interactions. This system allows us to solve optimization problems, which are reduced to finding the ground state for a set of Ising spins using the “quantum annealing” method — finding the global minimum of some objective function by means of quantum fluctuations (tunneling through potential barriers).

The capabilities of D-Wave One, its developers have recently demonstrated on the example of the problem of laying protein. Its essence is to find the most stable protein packaging with a specific amino acid sequence. A measure of the stability of a structure is called free energy, the global minimum of which the computer should have found.

For testing, scientists prepared an elementary sequence of six amino acids, with which the 128 qubit processor, cooled to 20 mK, successfully coped. In reality, quantum computers will have to solve much more complex problems, and D-Wave Systems specialists, realizing this, are trying to increase the reliability of calculations and release more powerful processor modifications.

[Iraclius]

The most cimmess is that a D-Wave One computer with an 128-qubit chipset, the cost of which, I recall, $ 11000000, performs only one task - discrete optimization.
Here's a cut in American.
Best of all, the activities of D-Wave are characterized as follows:

For example, Scott Aaronson, a professor at the Massachusetts Institute of Technology, believes that D-Wave has not yet been able to prove that her computer solves any problems faster than a regular computer, or that it is possible to enter the used 128 qubits in a state of quantum entanglement. If the qubits are not in an entangled state, then this is not a quantum computer.

[Iraclius]

I also want to recall the Japanese epic with the creation of computers of the 5 generation, which took 10 years (from 1982 to 1992 a year) and about half a billion dollars. The program, of course, ended in an equally epic failure. In this connection, science fiction writers sharply slowed down and AI began to disappear even from science fiction novels. Some scientists, within the framework of the theory of singularity, generally deny the possibility of creating thinking machines, claiming the uniqueness of human creativity and intellectual activity.

[queen]

There is always a Chinese switch on the power supply))))))))))))))

[Uruska]

All this will be tested on us. There are no volunteers!

[Iraclius]

The news is quite old, but this is the maximum of what can be found in the open press:

According to eWeek, IBM scientists have created a prototype of a cognitive processor. Its architecture is fundamentally different from the traditional one, and the method of information processing is based not on the "if-then" principle, but on learning ability, variable responses and the search for ways of optimal solutions, depending on the task at hand. The lead developer of this project, Dharmendra Modha, argues: "Future programs will need to expand the functionality of computers in order to work, and modern architecture will no longer be able to provide it. These processors are an important step in transforming a computer from a calculator into a learning system, and signal about the imminent appearance of a new generation of devices and applications that help in business, science and education. "

Crystals are grown from silicon, containing blocks that are built-in memory (artificial synapses), information processors (artificial neurons) and conductors (artificial axons). The working prototype of the cognitive processor has two 45-nanometer cores made according to the SOI-CMOS standard. One core contains 262 144 programmable synapses, the other contains 65 536 trained synapses. Now tests are being conducted in the tasks of pattern recognition, associative memory and classification. Attempts are also being made to use the new processor in navigation and artificial vision systems.

Despite the optimism of scientists, the application of such chips is very limited.

[Honory]

In my opinion, a very necessary program. Because it will be useful not only for the military.