The three questions that artificial intelligence will not be able to answer

36 תגובות

1. I guess the machines can think.
   They just think differently.

2. rival
   So here you are wrong. If you take a neural network that was trained to recognize cats, and teach it to recognize fish - then it will no longer know how to recognize cats. It will change the weights in the sensors calibrated for one purpose, for the other purpose.

   Yes - an adult learns a second language in a different area of ​​the brain. We know that for sure. In general, the subject of "areas" is foreign to artificial networks, but it is a fundamental part of the structure of the brain.

   If what you say were true, we would expect different people to have the same level of intelligence, or perhaps intelligence would be directly related to brain size. This is far from true.

   Do you think that our ability to recognize certain things depends on the connections between the neurons? Our brain is designed to divide the world into 5: people, animals, plants, tools and objects in nature. (There is a broad empirical basis for this). Can it be said that this is due to something in the connection of the neurons? On what basis can you claim that?

3. Miracles,

   Are you saying that in adults a second language is stored in a different place in the brain, and in children both languages ​​are stored in the same place?

   In any case, I do not agree with you that there are mechanisms in the brain that are above the basic network and that this is something that we really do not understand... In my opinion, it is simply a matter of how the neural network is organized from the inside, for example, when you train a neural network to recognize images of cats and dogs, will they both store in the network in the same place? Or in different places? And if you first trained a neural network for a long time to recognize only cats, and after it became an expert in recognizing cats, you suddenly started teaching it to recognize fish as well, will the fish and cats keep in the same place in the network or not?

   By the way, we did something similar with Deep-Mind's neural network, the engineers there discovered that if the network trains only against itself, it reaches much higher achievements than if it first trained on a lot of human games, and only then started playing against itself. It reminds a little of the matter of learning languages, doesn't it?

   As for the tapir, it depends on the patterns that your brain knows and has learned to recognize... if a dog passes by and that local points to it and tells you "Bogitmo" you will probably think it is "dog" in the local language, but if he points to the dog and tells you "Max", "Rexy" Or "Mookie" you will probably think is the name of this particular dog, it all depends on the circumstances and the things the neural network already knows.

4. rival
   The problem is not the amount of neurons - there is no such thing as an "unused" neuron. All brain cells are connected all the time. What happens is that a second language in an adult is stored elsewhere in the brain. I mean - there are mechanisms in the brain that are above the basic network, and this is something we really don't understand today, so there is no way that someone has implemented it correctly in an artificial network.

   What you said about the tapir is true. But why is this true? How do you know he didn't mean the name of the tapir (if it was my dog ​​I would say "Becky", not a dog). Why doesn't the word mean "food", or "animal", or "runner"? Why the species of the animal? The reason is (probably) that we have a learning mechanism in the brain, an innate mechanism. Otherwise, how can a baby learn names of objects? How does he distinguish between a "dog" and a "bucky"?
   That is - we have layers in the brain above the network of neurons that constitutes the infrastructure. Other animals do not have this layer, but we know nothing about this mechanism (at the level of neurobiology, psychologists have more understanding on the subject).
   Until we understand these mechanisms, we are very far from building an intelligence like that of man.

5. Miracles,

   Common sense tells me of course that the meaning of the word he said is "tapir".

   Regarding two languages ​​that a child remembers both and an adult forgets the previous language actually sounds very logical to me. A child's brain has a much larger amount of neurons, so the brain can afford to use some of them for one language and another group of neurons for the other language.

   In an adult, on the other hand, the brain cannot be all generous because the amount of resources is smaller... therefore it is forced to allocate some of the neurons that were used for the first language for the second language.

6. rival
   I think Google was used to translate, but I have no way of knowing.
   We are very far from understanding how the brain translates languages. We know that there is a fundamental difference between a baby who learned two languages ​​and an adult who learned a second language (one of the differences, but not the only one, is that in adulthood there is a tendency to forget a second language, if it was learned late).
   More than that - we distinguish between two types of second language: pidgin and creole. Broadly speaking, the first type is what adults who move to a foreign country learn, and the second type is the language that the children of these people learn from their parents (the differences are in the grammar of the languages).

   The way humans acquire a language is not similar to the way it is taught on computers. We have structures in the brain whose job it is to process language, and we have no idea how they work.

   We don't have much idea how the brain processes language, certainly not the level of neurons. We don't even have an idea of ​​how many neuron spindles there are in the brain (they claim around 10,000). You can also say that our brain and a computer work in the same way, because both have electrical signals. It doesn't mean anything 🙂 Today we have no practical way to study the human brain - MRI machines do not see neurons, and since they connect to a neuron - this does not include the synapses (there are sometimes tens of thousands per single cell).

   Let me ask you a question. You walk in the Amazon with a local who doesn't speak English. At a certain moment a tapir crosses the road - the local points to it and says "galagai". What do you think this word means?

7. Miracles,

   The question is whether the translation you are talking about is done using a neural network or using other methods. According to what I understand, the translation using a neural network is already starting to become much more natural and accurate than what we have seen so far:

   https://www.tgspot.co.il/google-translate-is-now-able-to-translate-even-more-languages/

   And by the way, I wouldn't be in a hurry to say that the network "doesn't understand" what it is translating, overall, the way it translates is quite similar to what is done in our brains, at least at the basic level.

8. rival
   Mainly on one site that centers news in the VR world. The articles are translated into English, but I don't know from which languages. I notice that the articles (in part) are translated because there are funny mistakes. Once it was written "expensive reader", and once it was written "monkey paper". I know that the last one is a translation from Hebrew (copy paper....).

9. Miracles,

   Just curious, where do you get to read a lot of computer translated articles? Is this a certain website? From which language to which language is the translation?

10. There is no doubt that connecting a computer to a human brain is already here. The idea that a person who does not understand the machine will connect to a computer and play Mozart with emotion even if he has never touched a piano... The next question is whether it is possible to do this by remote control with radio waves for example....
    From here to the ruler who controls rulers, that is, the king of kings: {Putin as the most powerful man in the world} the road is short..
    Puppet rule and nothing to do when the computer man takes command...

11. rival
    Translation programs do not understand the language. They do little more than a dictionary search. I get to read many articles translated by computer, and it is clear beyond any doubt that a person did not do the translation. I once saw a western where someone said (that is, in a written translation) "He ran as if he saw an air-to-air missile"... Go explain to the computer that in the context of a western, a sidewinder is a species of venomous desert snake....

12. It is quite clear that everything (!) claimed in the article and attributed to the Chinese writer Hao Jingfang, is simply not true.
    Explanation: All possible questions can be divided into two groups:
    One - knowledge and calculation questions, such as the first question.
    There are correct answers to these questions.
    If the artificial intelligence is "smart" enough, then it will know the correct answer.
    If she does not know the correct answer to the question that will be asked, then she will answer something like this:
    I am not familiar with the subject (for example, "Chinese food" or "foods originating in America"), I would be happy to answer the question after I study it. Because the definition of artificial intelligence is the ability to learn.
    It is likely that the artificial intelligence, after learning the subject, will be able to provide an even more comprehensive and inclusive answer than the one provided by Ms. Hao Gingfang.
    The question is, how long will it take the artificial intelligence to learn the subject? It already depends on the speed of the computer and the databases to which it is connected. It doesn't seem like the process will take more than a few seconds.
    The second group are questions that do not have a correct answer.
    The question then, refers to the "discretion" and "personal preference" of the artificial intelligence, so there are two options:
    One: that from the beginning it was given certain character traits by the manufacturer. Then she will choose the answer according to the character traits according to which she was created.
    The second: if she does not have a clear preference due to her character traits, she will choose one of the options randomly (lottery).
    This happens many times in our lives as well, those of us who have "natural intelligence".

13. Miracles,

    Language probably plays an important role in thinking, intelligent problem solving, and self-awareness, but why do you think neural networks can't learn language? I know that today there are quite a few neural networks that specialize in learning languages, for example neural networks that are used to translate between one language and another.

14. Yosef
    "Such a network with a depth of 101 layers compared to 1 is very close to imitating the human thinking behavior and therefore as an expert system it specializes best in the profession that it is required to acquire"

    A structured network and an expert system are almost opposites. And both are not necessarily related to the human way of thinking. I will ask you a simple question - do you need a language to think? The answer is - that we really don't know the answer, and this is discussed a lot in both philosophy and science.
    [An expert system describes a completely different architecture than a design network].

15. Regarding proximity to human behavior, there will be those who disagree with me. Let's say very effective behavior.

16. In this field, we see a Chinese hegemony albeit of doctors from the USA. For example Professor Fei Fei Ling from Stanford.
    Also in articles. marks the rise of China. We also see child professors (26) who come from Africa, China and India
    To Oxford, MIT, Stanford, Toronto and more.

17. In 2012, a number of factors combined that made possible the transformation of a convolutional neural network
    To great effect. The introduction of the NVIDIA graphics processor (GPU) which greatly accelerated the calculations in neural networks. Introducing regularization methods that direct the layered network to train correctly. The introduction of Google's residual networks, which made it possible to train networks over 5 layers in a cascade up to 101 layers and even more.
    How is a multilayer neural network different from a normal one? Normal is single layer. Revoda is multi-layered and allows for abstraction
    of the mixed information for a variety of features. In this way, if one network is wrong, the second is not wrong or the third. Experiments also show that the information abstraction works between the layers. Well such a network with a depth of 101 layers compared to 1 is very close to imitating the human thinking behavior and therefore as an expert system it specializes best in the profession it is required to acquire. As a result of all these layered networks that have been trained, almost nothing can confuse them and they beat a human competitor easily. In the breakthrough of 2012 the training took 3 months. Today it takes 3 hours.
    Another person named Professor Ian Judfellow (30 years old) managed to make two networks compete with each other and in this way train each other instead of multiplying the experimental sample. His networks are Google networks called adversary network.

18. Yosef
    Every day the situation is not the same as it was yesterday. What happened in 2012?

19. I don't intend to enter a chain of comments and not out of arrogance.
    The situation today after 2012 is not the previous situation.

    There was nothing in the achievements, in the understanding, in the power of calculation.
    The current target of artificial intelligence is really an expert doctor to decipher MRI. You hear about an IBM project on the matter. We do not understand at the human level yet. Even you see that within a decade you will have an autonomous vehicle and many similar things.

20. Yosef
    I've been messing around with simulating a worm's brain. In the worm's brain - 302 neurons and about 6400 synapses. One of the things we know is that it is impossible to simulate a brain without the muscles, so we simulate the entire muscle system of the worm - all 95 cells :).

    The rumor talks about six-year-old wisdom? The reality is that we are nowhere near simulating this worm! The difference is that we are not trying to sell something….

    I worked for many years in two very different fields, but both are relevant to the discussion. The first is simulators of fighter planes, in particular - simulation of a tactical arena. The idea is to simulate to the trainee (or the group of trainees) structures of enemy planes that will behave in a natural way. What we learned is something important: the pilot needs to be simulated on many levels. The low level is hand-eye coordination (putting a cannon sight for example). Above that is basic flight - straight and horizontal flight, coordinated turn and so on. Above that - aerobatic exercises (loops, barrel rolls, Immelman...). On this are basic combat maneuvers - countering a cannon, sewing balances, tilt reversals, etc. And in order not to get bored - we will skip to the higher levels such as defensive patrol, offensive interception - and continue to develop combat theory according to the type of aircraft and the culture in that country.
    All these are aspects of Bina - but it is clear that each level needs a different implementation (and I did not even mention the fatigue during the flight, due to lack of fuel, loss of a spouse, a member of the squadron falling into captivity, or the political situation...).

    The second field is ... brain surgery. We built software that builds a XNUMXD model from various scans (MRI, CTI, DTI, fMRI, PET and more). I worked closely with brain surgeons to learn what our model looked like, compared to the brain itself. The resolution of these scanners is a few tenths of a mm, and the resolution of the information flow - let's say that every doctor sees something different there....

    We have a great deal of knowledge about the brain, but in terms of percentages - we know nothing. Before AI, everyone was excited about Big Data. In two years the AI ​​will become something else. The main thing is to sell...

21. If you believe that it will happen in 50 years and sleep the better, and I believe that it may happen within a decade, the fundamental difference is not that great.

22. Before 2012 there was no hardware available for such calculations. Today I bought my son a platform for computer games for about NIS 5000 that allows me to perform such experiments. In the laboratory there is a server costing about NIS 30,000 that speeds up such experiments from hours to minutes. Today there are software tools that allow me to change the structure (architecture) immediately. In the past you had to program them manually like assembler. We are in the era of artificial intelligences that are not human - I agree with you. I don't think the distance is as big as you think, it is substantial but it will be crossed.
    You also need to plot the quality discoveries on the timeline. For example, networks that compete with each other according to a zero-sum game and train each other in the way of Professor Ian Judfellow (30 years old), or the information bottleneck theory as a key to understanding deep learning by Professor Naftali Tashvi (about 60 years old with a young soul), or the theory of Geoffrey Hinton.
    The points run much faster now than before 2012. Then there was a time constant of a decade. He is now two years old.

23. The debate in my opinion is unnecessary. I simply believe, not blind faith, that some years will pass and they will catch up with us. And it's not that far.
    If you deal with artificial intelligence, you see with your own eyes and are not fed by a third party.
    Artificial intelligence is to a large extent an imitation of human intelligence. The fact that it is accessible to millions of researchers worldwide for laboratory research, and that there was a breakthrough that we all feel in 2012, means that a huge number of laboratories are investigating a human intelligence-like system and can design it for experimentation at will. A new factor was created - multiple experiments, due to the availability of laboratories. I have a lab at home because Google gave me tools.

    Second, let's put aside for a moment the aversion to the fact that they can simulate us, there are researchers who are trying to understand:
    A. How deep learning works
    B. How consciousness develops and
    They reach very similar results. and succeed in quantifying concepts such as consciousness, internalization.

    The rumor alone speaks of 6-year-old wisdom in Google's labs. And two years ago a two-year-old wisdom.
    In your brain and mine separately 15 billion neurons, we are close to billions of artificial neurons at hand. It is clear to both of us that the quantity alone does not do the job. There are also qualities here.
    There are metastructures (called architectures) of neural networks that perform the "thinking" (the quotation marks are forced in my opinion) in a much more efficient way than other structures. Today they know how to plan the quality as well. When 2 such programs talk to each other in a secret language, why do you think it is so different from you. To me it is not significantly different.

24. A little curiosity. I noticed that since I started dealing with elite artificial intelligence (CNN) the antivirus sends me a warning about its elimination of f8 files that reach me from the outside - I didn't recognize through whom. Today during my self-study, I noticed
    There is a research group that deals with especially deep networks called VGG. Their programs end in f8.
    These cookies may be artificial intelligence machines, software that may be sent by a super company that offers, for example, free mail,
    To understand what a person is doing. Fortunately, I purchased a powerful antivirus (NIS 40) and it still eliminates such threats.
    There are no free meals. It is not delusional. It is part of the present.

25. Yosef
    An artificial neural network is not artificial intelligence, for the simple reason that we do not know what intelligence is.

    The first point is that the correct term is machine learning, and this is not the way human intelligence works. What we do learn this way has nothing to do with intelligence (music for example).

    Second point - all kinds of futurists throw us amazing predictions. Ray Kurzweil promised us a singularity in 20 years, but keeps restarting the countdown. At least in one thing he was wrong - the rate of increase in the separation ability of the MRI devices is very low. To mention - Kurzweil claimed that in 2010 the computer will have the same computing power as the human brain.

    Third point - the techniques of AI (machine learning, structural networks, pattern recognition) have been known for decades.

    Fourth point - the lack of proportion. True - today's computer is the world champion in chess. But let's do a little exercise: let's change the rules so that a knight can only move two squares in a certain direction, but the next move must be to the right. How long will it take a person to adapt to the change, and how long to "deep blue"?
    Watson should know how to diagnose cancer today, right? So no - the computer is still learning what types of cancer exist... and if it does have some ability to recommend treatment - it's just a search in a table, built by doctors.

    I certainly agree that computers and neural networks do amazing things, but they are still less intelligent than a XNUMX year old baby.

    I personally think that the world of horticultural money is the one that will cause a revolution in the world soon. And the world of autonomous vehicles will bring about a smaller revolution, at a later stage.

26. Something that is already seen, let's say, before the age of self-aware machines like us, is a change in world order.
    From the moment I was fully exposed to technology, because I understand, I will not change my mind. They will catch us. Changing world orders. For example Intel made its second decision late. The first was not to enter the cell phone. Now NVIDIA leads the hardware for artificial intelligence. The super companies Google, Facebook, Tesla, Amazon - dominate scientific R&D.
    Google gives scientists the world's best artificial intelligence development environment for free. Environments for which 2 million NIS are paid are inferior to Google's free environment. Once made available to millions of human researchers, the development of understanding how we think increases exponentially. The researchers are a self-learning network. Apart. Similar to Ian Judfellow's adversary network.
    Such a network teaches itself, even in the minority of external study attribution. and teaches fast. Some companies have already woken up to the day when they have no future. They still exist. I too can be harmed by these economic changes. I don't think of it that way.

27. The way I would approach the problem is NLP - natural language, as well as a machine that identifies the connections between different content areas.
    What was limiting the amount of objects (called classes in artificial intelligence language) was the amount of computation required for linked object detection methods. The problem that the woman describes is similar in principle to the problem of identifying a dog + a person as distinct from a person.
    Although at a much higher level of complexity. Google doesn't tell everything they research. I take a risk and estimate that they are striving for artificial intelligence with abilities close to humanity, and of course the existence and evolution of understanding the human brain, from the moment that intelligence
    This has been exposed to millions of researchers (for Chinese, Indians and Africans it is an entrance ticket to the elite of the academic world and this is what you see there) has a natural experimental environment. We are a network ourselves through GOOGLE that will pilot artificial intelligence at a rate that biological evolution will not come close to. Names like Joshua Bengio, Ian Judfellow, Aaron Courville, Jan LeCun, Naftali Tashvi, Geoffrey Hinton. One should be a Nobel laureate for artificial intelligence because it is a field where the mathematical achievements are enormous and its influence on our lives is great.

28. As someone who is familiar with deep learning and is beginning to develop it at a high level, I estimate that by designing a suitable machine, the questions can be answered already today. Although where the answer is personal, she will answer a personal answer. Regarding understanding themselves: a. we don't know
    At lower levels of consciousness than us, animals are at least self-aware. plants? Mushroom networks according to many researchers yes. Let's agree that animals at the level of mammals are self-aware.

    B. Regarding artificial intelligence's self-awareness, 2 researchers are already carrying out work that shows that we are on our way there: Professor Naftali Tashvi from the Hebrew University of Jerusalem, Professor Giolo Tononi. There are different levels of consciousness. This is measured by the amount of information
    P*logP. The understanding may not yet be at our level, but it exists. Maybe to a much lesser extent, not sure at all.
    They have an opinion. In Facebook, two intelligent machines of the type described here, developed a secret language. It was still possible to turn off the electricity to them and that's what they did. At Tesla, such a machine for the application of an autonomous vehicle has begun to perform an optimization that endangers a passenger.
    Elon Musk concluded that there is a risk to the human race in general. He is currently developing an intelligence that is friendly to humans.
    Bill Gates and Stephen Hawking are very anxious for the future of the human race. I say paradox. I develop artificial intelligence machines even though I believe they will replace us. why? Because if I don't open the development will not stop. For the simple reason that in terms of profession - it is a good profession. Because it is self-admiring to produce a thinking machine.

29. Regarding the first question, if a programmer cannot answer this in 20 years, she will be able to in 30 (insert alternative dates according to your level of optimism). Beyond that, the correct answer is of course that the menu will not change substantially, because America would have been discovered in one way or another within 20 to 200 years from Columbus's voyage, and we are long enough after that for everything to be assimilated into the alternative history as well.
    Regarding the other questions, you will define artificial intelligence and then it will be possible to discuss if and when you can answer a specific question, and even give a hug to the questioner.

30. It seems that at least the first question has an algorithmic solution. First search: IF NOT, search words Columbus, discovery of the new world, types of food. The computer checks what types of food Columbus brought from America. Second search: Dish ingredients from Sichuan. Third search: Compare the results of the first search to the second. Answer: Dishes that have an ingredient that is equal between the first and second search.

31. The example is not clear at first. Let's start with the fact that not every person would have been able to answer the question
    "If Columbus had not discovered the New World, what dishes in Chinese cuisine would have been affected? "
    Maybe most people wouldn't know what the answer is
    It seems that this is exactly what systems like IBM's Watson, which won the Jeopardy trivia game, will do better than us, if not now, then in the coming years in all matters of trivia and the connectivity of the information these systems will leave us behind, in the future there is more entry into the differences of understanding versus simulation of understanding where we need " To understand" what does a person who understands mean and what is the difference between that and retrieving data from a computer. Is there a fundamental difference in understanding and what it is,
    Regarding apparent consciousness, it seems that even the most powerful supercomputers in the world have no consciousness at all, less than an amoeba, zero self-consciousness, the computer has no urges, no inner world, in this it is a machine for everything, playing X-Derix Mix or running a simulation of the universe or playing Go-All, doing nothing is equivalent to it It "excites" him in the same way, zero enthusiasm, zero urges, zero desires is something very different from a person, although from this you can tell that certain aspects of intelligence that characterize us so much can be done much better by a system with zero self-awareness,
    The Turing test is problematic because a system can deceive us like a set, but Panamit is not a representation of a person,
    Another topic that is of interest is whether the GAI system without internal drives and a consciousness that builds its own learning over time will it be able to "deceive" us, there is a feeling that it will not last long because the additional layers that will be built will not sit on
    The substrate of the inner feeling and consciousness systems which is like a kind of feedback equation that would be missing
    So that there will be no obstacle to building a completely different system without the same direction paths that we have, even with humans we get a huge variation here the variation will be even greater, the result will not be human but something else.

32. There are two confusing inaccuracies here. The first is the difference between consciousness and intelligence and thinking. Gingfang asks about thinking and infers from this to intelligence and consciousness.

    The second inaccuracy is the use of the word algorithm. There is a precise definition for an algorithm (thanks to Mr. Turing and others) - what the human brain performs, and also certain "smart systems" do not meet this definition. In particular - an artificial neural network does not run algorithms.

33. It is clear that the speed of the processor and the size of the transistors do not affect the intelligence of the computer, but a combination of extremely fast processing power with appropriate algorithms does. There is nothing new here that the author teaches any average science fiction fan. The progress over the years is with both although the software advances more slowly than the hardware

34. The old fallacy of assuming what the realization is

This site uses Akismat to prevent spam messages. Click here to learn how your response data is processed.