Rebooting AI - postulates

In recent weeks I've been forced to reformulate and distill my views on AI. After my winter post went viral many people contacted me over email and on twitter with many good suggestions. Since there is now more attention to what I have to offer,  I decided to write down in a condensed form what I think is wrong with our approach to AI and what could we fix. Here are my 10 points:

  1. We are trapped by Turing's definition of intelligence. In his famous formulation Turing confined intelligence as a solution to a verbal game played against humans. This in particular sets intelligence as a (1) solution to a game, and (2) puts human in the judgement position. This definition is extremely deceptive and has not served the field well. Dogs, monkeys, elephants and even rodents are very intelligent creatures but are not verbal and hence would fail the Turing test.
  2. The central problem of AI is Moravec's Paradox. It is vastly more stark today than it was when it was originally formulated in 1988 and the fact we've done so little to address it over those 30 years is embarrassing. The central thesis of the paradox is that apparently simplest reality is more complex than the most complex game. We are obsessed with superhuman performance in games (and other restricted and well defined universes of discourse such as datasets) as an indicator of intelligence, a position coherent with the Turing test. We completely ignore the fact that it is the reality itself rather than a committee of humans that makes ultimate judgements on the intelligence of actors.
  3. Our models may even work, but often for the wrong reason. I've elaborated on that in my other posts [1], [2], [3], [4], deep learning comes in as a handy example. We apparently solved object recognition, but numerous studies show that the reasons why deep nets recognize objects are vastly different from the reasons why humans detect object. For a person concerned with fooling humans in the spirit of the Turing test this may not be important. For a person who is concerned with the ability of an artificial agent to deal with unexpected (out of domain) reality this is of central importance.
  4. Reality is not  a game. If anything, it is an infinite collection of games with ever changing rules. Anytime some major development happens, the rules of the game are being rewritten and all the players need to adjust or they die. Intelligence is a mechanism evolved to allow agents to solve this problem. Since intelligence is a mechanism to help us play the "game with ever changing rules", it is no wonder that as a side effect it allows us to play actual games with fixed set of rules. That said the opposite is not true: building machines that exceed our capabilities in playing fixed-rule games tells us close to nothing about how to build a system that could play a "game with ever changing rules".
  5. There are certain rules in physical reality that don't change - these are the laws of physics. We have verbalized them and used them to make predictions that allowed us to build the civilization. But every organism on this planet masters these rules non verbally in order to be able to behave in the physical environment. A child knows the apple will fall from the tree way before it learns about Newtonian dynamics.
  6. Our statistical models for vision are vastly insufficient as they only rely on frozen in time appearance of things and human-assigned abstract label. A deep-net can see millions of images of apples on trees and will never figure out the law of gravity (and many other things which are absolutely obvious to us).
  7. The hard thing about common sense is that it is so obvious to us, it is very hard to even verbalize and consequently label in the data. We have a giant blindspot that covers everything which is "obvious". Consequently we can't teach computers common sense, not only because it would likely be impractical, but more fundamentally because we don't even realize what is it. We don't realize until our robot does something extremely stupid and only then an eureka moment arrises - "oh it does not understand that ... [put any obvious fact of choice here] ...".
  8. If we want to address Moravec's paradox [which in my opinion should be the focal point of any serious AI effort today] we somehow need to mimic the ability of organisms to learn purely from observing the world, without the need of labels. A promising idea towards achieving this goal is to build systems that make temporal prediction of future events and learn by comparing the actual development with their prediction. Numerous experiments suggest that this is indeed what is going on in biological brains and it makes a lot of sense from numerous perspectives, as these systems, among other things would have to learn the laws of physics (as they appear observed by the agent, aka. folks physics). The predictive vision model is a step in that direction but certainly not the last step.
  9. We desperately need to frame the quality of "intelligence" outside of the Turing's definition. Promising ideas arise from non-equilibrium thermodynamics and are consistent with the predictive hypothesis. We need that because we need to build intelligent agents that will certainly fail the Turing test (since they will not exhibit verbal intelligence) and yet we need a framework to measure our progress.
  10. Almost all that we do today and call AI is some form of automation of things that can be verbalized. In many areas this may work, but is really not very different from putting Excel in place of a paper spreadsheet to help accountants. The area which is (and always was) problematic is autonomy. Autonomy is not automation. Autonomy means a lot more than just automation, and it means a whole lot more if it is autonomy that is required to be safer than humans, as in self driving cars. Autonomy should almost be synonymous with broadly defined intelligence as it assumes ability to deal with unexpected, untrained, proverbial unknown unknowns.

These are the core points I'd like to convey. They have various nuances, hence why I write this blog. However certainly if you acknowledge these points, we are pretty much on the same page. There are other numerous details which are heavily debated which I don't think are essential but for completeness let me express my views on a few of those:

  1. Innate or learned? Certainly there are organisms with innate capabilities and certainly there are things we learn. This is however an implementation related question and I don't think it has a definite answer. In our future development I'm sure we will use the combination of both.
  2. Learned features or hand crafted features? This is a related question. My broad view is that vast majority of aspects of the "cortical computation" will be learned, that is in the context of AI and autonomy (but that does not mean we can't handcraft something if it proves to be useful and otherwise hard to learn for some reason). There are also huge pieces of brain that are most likely pre-wired.  In more specific application of automation, things can go both way. There are cases in which learned features are clearly superior than hand crafted ones (the whole sales pitch of deep learning), but there are numerous applications where carefully handcrafted and developed features are absolutely, unquestionably superior to any learned stuff. In general I think it is a false dichotomy.
  3. Spiking, continuous, digital or analog, maybe quantum? I don't have an extremely strong position on that, each has advantages and disadvantages. Digital is simple, deterministic and readily available. Analog is hard to control but uses far less power. Same with spiking though that has the added benefit of being closer to biology which may suggest that for some reason this is the better solution. Quantum? I'm not sure there is any strong evidence for the necessity of quantum computation in solving intelligence, though we may find out it is necessary as we go. These are all questions about "how?". My main interest is in the question of "what?".

Since I want to keep it short (it is already too long) I'll stop here. Feel free to give me feedback in comments, emails and on twitter.

Comments

comments

10 thoughts on “Rebooting AI - postulates

  1. I believe #8, "build systems that make temporal prediction of future events and learn by comparing the actual development with their prediction", is in fact the correct definition of intelligence. I take this framing from Jeff Hawkin's 2004 book On Intelligence, but it is not unique to Hawkins of course. So this definition should replace the turing test. And as far as I can tell, that's pretty much what's happened in practice.

    Note that I would include predictions of how your own actions will impact the future is part of this baseline definition. So it's predictions of what will happen, and how my possible actions will change what will happen.

    Enjoy your post. Thanks!

  2. This is also true for socially intelligent machines, which must model both themselves and others to predict the future and compare it with their model's predicted future.

  3. Very refreshing view, thanks for not going with the crowd here. Your #8 made me think of the work on reinforcement learning, but you don't cite this field. Is it because it is primarily applied to closed games currently?

    1. Yes, RL is approaching the problem completely from the back. We should stop focusing on very specific goal oriented strategies. It is an important part, but probably takes maybe 5% of the brain. What we are lacking is the simple, embodied perception in the real physical world. We don't generally have a concept for that, and we don't generally even recognize this bit is missing. It is the basic rules of interaction between a cognitive agent and physical reality that we are totally missing, and Moravec's paradox is the result of this. So yeah, I'm not particularly excited with RL, particularly when applied to closed games, stuff being done to death right now.

  4. "2. [..] We completely ignore the fact that it is the reality itself rather than a committee of humans that makes ultimate judgements on the intelligence of actors."

    Not true. "Intelligence" is a word and words are defined by humans. Thus, whether someone is intelligent or not is defined by the manner in which humans define intelligence -- not by something external to humans.

    What reality can determine is the degree to which a belief is true. But someone whose beliefs are true is not necessarily intelligent. As Plato said long time ago, knowledge is justified true belief which means that a person's beliefs must be justified, i.e. properly formed, before that person can be considered intelligent. A person who forms their beliefs by flipping a coin cannot be considered intelligent even if every single one of their randomly generated beliefs happens to be true.

    What we call intelligence is a trait that has evolved within a specific type of environment -- and that means within a specific box (i.e. a set of constraints on what is possible.) It does not work, it cannot work, outside of this box. There is no method of forming beliefs that can work within ANY kind of imaginable environment. For every method of forming beliefs, there's an environment within which it performs worse than other methods.

    "1. [..] Dogs, monkeys, elephants and even rodents are very intelligent creatures but are not verbal and hence would fail the Turing test."

    This, on the other hand, is true. A lot of people mistakenly believe that intelligence is something that is fundamentally verbal. That's not true. The verbal part is the superficial part.

    "3. [..] For a person who is concerned with the ability of an artificial agent to deal with unexpected (out of domain) reality this is of central importance."

    There can never be such a thing. What you can do is create an agent that can deal with a relatively broader range of situations but you can never create an agent that can deal with every kind of situation imaginable (such as unexpected, out of domain, situations.) Building a machine means creating a plan, a set of rules, that your machine should obey without any questions. When it doesn't do so, we say the machine is broken or we say we are not responsible for its behavior. Machines have a backbone and they can't adjust to any situation without deviating from their design.

    The real problem is that people create machines that can only deal with a narrow set of situations which then they try to apply to situations for which they were not designed. The solution is to clearly define the task your machine should perform so that you can better anticipate how well your machine will perform (and not be so surprised when it performs poorly.)

    "4. Reality is not a game. If anything, it is an infinite collection of games with ever changing rules. Anytime some major development happens, the rules of the game are being rewritten and all the players need to adjust or they die. Intelligence is a mechanism evolved to allow agents to solve this problem. Since intelligence is a mechanism to help us play the "game with ever changing rules", it is no wonder that as a side effect it allows us to play actual games with fixed set of rules. That said the opposite is not true: building machines that exceed our capabilities in playing fixed-rule games tells us close to nothing about how to build a system that could play a "game with ever changing rules"."

    Intelligence has evolved within relatively stable environments. No stability, no intelligence. Stability (or regularity) means nothing other than constraints, limits, boundaries, rules. While I can sort of, kind of, agree that reality is not a game, I cannot agree that intelligence is something that allows us to win at any kind of game. Intelligence is not a superpower.

    1. Regarding your first point. "Energy" is just a word. And yet we believe there is some objective quantity this word describes. "Energy" has no ultimate definition, but we can measure it, we can model it, we can talk about it and use it. Sam for "intelligence". As for the other things, yes I agree, there are limits. Indeed intelligence cannot deal with any situation, otherwise nobody would ever get trapped in a situation without exit. But it is extremely robust and it handles changing reality. You claim that intelligence requires stability. To some degree yes. But there is very little stability between say XV century and now. Almost everything in terms of human environment has changed and yet we deal with it without much problem. You can take a person, send him to a different country and within a few days they will accommodate just fine. We currently have no technology that could do the same thing. My take on it, is we don't really understand the "machine learning for long tails".

      1. If I understand you correctly, what you're saying regarding my first point is that we can understand the meaning of a word without being able to describe it in terms of other words. I agree with this. My point is merely that words, as well as their meaning i.e. how they are used, is determined by us. We invent words and we determine what they mean. This process is not necessarily a verbal one. (In fact, I think that in most cases we establish the meaning of words through actions, i.e. non-verbally, rather than through other words.)

        Understanding what intelligence is boils down to understanding how we use the word "intelligence" in practice. That's all there is to it.

        I do, however, agree that Turing's definition of intelligence is inadequate. Intelligence is about making predictions in a specific way (not necessarily accurate, since human intelligence is fallible), not about having a realistic conversation with a human being using a natural language. That would be deep/broad versus shallow/narrow inteligence. If you know how to make predictions using any kind of data available to you, it is straightforward (albeit expensive) to figure out how to have a realistic conversation with a human (all you need is adequate experience i.e. data.) The opposite, however, is not true. In both cases, we're dealing with a boxed intelligence, the only difference being the size of the box. Broad intelligence works within a huge box: it's extremely adaptable but not completely adaptable. Intelligence is a type of belief-formation that is based on existing beliefs such as past observation. It is based on an assumption that the future will be in certain relation to the past; specifically, that the future will mimic the past. Of course, this assumption isn't completely true, so we have make up for it by making as many direct observations about the world as possible. And it imposes a limit. You can't operate within any environment that isn't sufficiently stable.

        If you agree with this then we're on the same page and I'm merely being confused by minor imprecisions, i.e. noise, in your article.

        1. I think we are roughly on the same page, though the subject is complex and indeed it is hard to express it concisely. I think we lack the proper vocabulary to be unambiguous and precise about these matters, and I don't see much effort in the so called AI community to make this language more precise, which is partially what frustrates me.

          I do think however there is a qualitative (not just quantitative) difference between "intelligence" boxed in a box that is a (finite) formal system defined by a human and "intelligence" boxed by a (likely) non formal (i.e. not expressible by a finite set of axioms) system that we call reality. The first type we appear to be able to solve. The second probably does not have an ultimate solution (i.e. human intelligence is far from perfect I suppose and it is even hard to image what would a perfect intelligence look like - potentially an all knowing oracle or something), but whatever solutions we have in that domain simply suck compared to even human infants.

Leave a Reply