To Build Truly Intelligent Machines, Teach Them Cause and Effect (2018)
quantamagazine.org
quantamagazine.org
Working on common sense, defined as predicting what happens next from observations of the current state, is a classic AI problem on which little progress has been made. I used to remark that most of life is avoiding big mistakes in the next 30 seconds. If you can't do that, life will go very badly. Solving that problem is "common sense". It's not an abstraction.
The other classic problem where the field is stuck is robotic manipulation in unstructured situations. McCarthy once thought, in the 1960s, that it was a summer project to do that. He wanted a robot to assemble a Heathkit TV set kit. No way. (The TV set kit was actually purchased, sat around for years, and finally somebody assembled it and put it in a student lounge at Stanford.) 50 years later, unstructured manipulation still works very badly. Watch the DARPA Humanoid Challenge or the DARPA Manipulation Challenge videos from a few years ago.
Great PhD thesis topics for really good people. High-risk; you'll probably fail. Succeed, even partially, and you have a good career ahead.
When we think-about-thinking, or talk-about-thinking, we do so in the language of language, which quickly leads to logic. And that leads us to think that the logic is the thinking. But in fact it's a rare, specialized mode of thought. The primary mode of thought -- the one that keeps us from making big mistakes for a half-minute at a time -- is that irrational one that's very easy to fool if you put effort into it, but which actually gets it right for most of reality (which isn't, generally, trying to trick you).
Yes. Language is not thinking. Language is I/O.
The easiest argument against it being: have you even struggled to put a feeling into words? The answer, of course, being yes.
You might literally interpret this as "is internal monologue a form of thought" and the answer is clearly yes, on sum logical level, it's brain activity.
You might literally interpret this as "is internal monologue the ONLY form of thought you have" and the answer seems just as clearly no and the question seems simplistic.
But I think people who say that usually mean "is internal monologue the primary-in-some-way form of thought you have" and there the debate gets heated but if you reveal statement, you reveal the confusion is mostly in debating "what's primary in brain activity". But if you think about, lot of debates about human thought is about which part is primary in a fashion we can intuitively feel.
If language is how people think, then you should be able to convey your thoughts to pther people easily and perfectly by using language.
Obviously, that is true for only the most trivial thoughts, thus I personally conclude that language has hardly anything to with actual thinking.
Based on what I think goes around in my head when I think, I would say that language is one of the outputs of thinking. Which may happen as "internal speech",via mouth or pencil or keyboard.
Actually, I think the key thing is Language is not logic. People are great at simple logic statements and terrible at complex logic.
A line of SQL with 1-2 predicates can indeed seem "easy and language like" but a line of SQL with 5-6 predicates can be utterly opaque, a beast that many programmers would be happy to write an entire 1000-line C program to avoid.
I think a key failing of discussion of cause and effect, logic and so-forth is a failure to really come to terms with the multifaceted qualities of human language expression.
And I'd agree with other posters that the distinction between I/O and thought can be debatable.
After 3 dimensions it gets extremely hard and people quickly move to heuristics and short-cuts because most people (i.e. the vast majority including myself) can really only comprehend 3 dimensions of information.
What happens when a human hand grabs and moves an object? At least 5 dimensions that I can think of:
1. x axis 2. y axis 3. z axis (i.e. 3 axis's just to get to the object) 4. grip strength 5. rotation
Those 5 are the absolutely basic, and there are many others, speed, grip texture etc.
I know engineers have programmed robotic hands to grab stuff, but I cannot imagine how many dimensions there are in unstructured problems. That requires engineers to think through so many dimensions, which seems impossible to me.
How is "predicting what happens next" different from predicting, say, the next word in the sentence using modern DL models?
Still sounds super hard, but might be easier than the same (matching corresponding shapes) in 3D.
We'd expect some kind of AI-human parity in avoiding obstacles while driving in common sense speak as "don't hit that, or you'll have a wreck", but we don't really expect the car AI to see a bad collision between two other cars on a perpendicular road and call emergency services (as should be common sense for humans).
But if the cars involved in that collision have a detection system to automatically call 911 (any kind of OnStar variant), why should an AI concern itself with that knowing there is a system to handle that task? Would it be common sense for the AI to act as a parallel system and make sure the primary didn't fail to call 911? A human's common sense might be to act as if that system didn't exist because it might have failed and just call anyway (knowing that there's really no penalty for calling twice just to make sure)
If you compare these to Boston Dynamics, it is hard to watch. But to be fair, BD robots are remote controlled.
I doubt that you can create a mind that's similar to a human mind without the relevant elements that are took for granted when we think of a human being: senses, perception, pain, pleasure, fear, volition... a body! and the real-time feedback loop that connects us to our environment and our peers.
The same could be said about animals' minds. That's why it's still impossible to make even a mosquito brain. It's a question of texture. Making a decision for a human involves a complex cloud of subsystems working in unstable equilibrium, more of a boiling cauldron than an algorithmic checklist. When you're scared, you're not just thinking that somethind is dangerous and rather avoid it, you are feeling something very uncomfortable and you want to stop it.
What if you want to advance in creating some kind of simpler mind now when you still haven't the means to build a complete organism? That's an interesting problem. Would immersing programs in a virtual world be useful? Or would it be better to make robots face the real world directly? I believe that you need, as a minimum, a system that integrates sight with hearing and touching sensors, and some kind of incentive system.
After some results, maybe using machine learning, the emergent organization could be applied as a building block to more complex robots. Meanwhile, trying to teach machines some human capabilities will not lead to generalized IA, but to more of the same we have now, that it's very useful, just not quite qualifies for the label.
Yes! Almost all neural networks have no self-model and thus no self-awareness because they cannot perceive themselves. They only see the inputs. They do not see the result of their actions.
This makes developing a self-model impossible. They cannot develop an internal model of internal vs external causes. What their "boundary of influence" is. Differentiation between internal and external causes.
They are trained and then used, immutable, unlearning after training. Even if they could perceive their outputs during training and/or evaluation, they cannot perceive themselves otherwise, making it practically impossible to deduce by themselves what they even are. They can't inspect themselves.
The causal loop needs to be closed for all of this to happen.
Symbolic reasoning is basically learning a lot of "if-then" statements and chaining them to make inference. Causal reasoning consists of defining conditional dependencies of current state on past state, then extrapolating based on the encoded assumptions. It requires some notion of object relation both in a literal sense as well as subtler relationships. Regression techniques are being ham-fisted to fit these roles but the popular ML of today is still just pattern recognition and cannot be called "reasoning" per se.
I don't work directly in this space but I see it following closely the architecture of the human brain for a while before departing to more distilled forms of knowledge management structures.
Neural networks are capable of approximating any system to arbitrary precision.
There is a need to construct logically-deduced models which impose an inductive bias so that your regression methods are efficient. That's where reasoning comes in, and where automated reasoning methods should be useful.
You could probably argue that the cause/effect stuff is subsumed by one of these at a certain level of abstraction, but I think it makes sense to treat them as separate.
Related to the idea of "cause/effect" and possibly falling into the overall rubric of "intuitive metaphsyics" is some notion of the passage of time. That is, in human experience we link things as "causal" when they happen in a certain sequence, and within a certain degree of temporal proximity.
Eg, "I touched the hot burner then instantaneously felt excruciating pain" is an experience that we learn from. "I walked through the door and four days later I felt pain in my knee" probably is not.
Our machines probably also need baseline levels of some sort of intuitive versions of Temporal Logic and Modal Logic as well.
https://en.wikipedia.org/wiki/Metaphysics
https://en.wikipedia.org/wiki/Epistemology
(1) John took the water bottle out of the backpack so that it would be lighter.
(2) John took the water bottle out of the backpack so that it would be handy
What does it refer to in each sentence? It's very obvious that a machine that solves this must understand physics, have a rudimentary ontology about objects and human intuition and so on.
I think it's straight-up sad how little progress there has been on these very fundamental problems which articulate what common sense and intelligent agents are about.
I re-phrased the sentence as "John took the water bottle out of the backpack in order to make lighter|less heavy|handy the" etc.
The only way it would complete with something other than the water bottle was "John took the water bottle out of the backpack in order to lighten the " and it completed with 'load'
The most amusing was a task to fill in the blank with "John took the water bottle out of the backpack so that [MASK] would be lighter"
The model was 98% confident the blank should be 'it' facepalm
Interestingly enough when I change the fill in the blank sentence to:
"John took the water bottle out of the backpack so that the [MASK] would be lighter."
The result was: "23.9% liquid 13.4% water 7.7% contents 6.5% weight" and the last two are pretty close.
I ran these tests on https://demo.allennlp.org/.
I suppose that's cause-and-effect in a loose sense, but one doesn't have to view everything as C&E to get similar results. It seems more powerful to think of it as relationships instead of just C&E because then you get a more general relationship processing engine out of it instead of a single-purpose thing. Make C&E a sub-set of relationship processing. If the rest doesn't work, then you still have a C&E engine from it by shutting off some features.
I may be wrong (heck, I'm probably wrong), but I can't help but feel that you're abstracting things out too much. Yes, a "cause / effect relationship" IS-A "relationship", but sometimes the distinctions actually matter. I'd argue that a "cause/effect relationship" (and the associated reasoning) is markedly different in at least one important sense, and that is that it includes time in two senses: direction, and duration. There's a difference between knowing that Thing A and Thing B are "somehow" related, and knowing that "Doing Thing A causes Thing B to happen shortly afterwards" or whatever.
To may way of thinking, this is something like what Pearl is talking about here:
The key, he argues, is to replace reasoning by association with causal reasoning. Instead of the mere ability to correlate fever and malaria, machines need the capacity to reason that malaria causes fever.
That said, I do like your idea of trying to build the processing engine in such a way that you can turn features on and off, because I don't necessarily hold that "cause/effect" is the only kind of reasoning we need.
Re quote: "The key, he argues, is to replace reasoning by association with causal reasoning. Instead of the mere ability to correlate fever and malaria, machines need the capacity to reason that malaria causes fever."
Back to my original point, humans did just fine at "intelligence" before science came along. That's Step 3, we need step 2 first. Find correlations, and if it seems to be that part-1 happens before part-2, then the bot can infer something equivalent to a causal relationship. That may be the time element you are talking about. Perhaps it's a matter of interpreting what "causal" means. I see it as "finding relationships we can take advantage of to obtain our goals". Whether there is physics or chemistry behind a relationship is an unnecessary distraction (without other advances).
Re: That said, I do like your idea of trying to build the processing engine in such a way that you can turn features on and off, because I don't necessarily hold that "cause/effect" is the only kind of reasoning we need.
So we kind of agree. If it turns out I'm wrong and that we can make most of the bot work right with just causal relationships, then factor out the general purpose "graph processor" for efficiency and make it causal-centric.
Early childhood studies should be able to answer the question as to whether humans are wired to expect causal relationships. My understanding is that all children do develop that expectation, as do animals.
1. There is a time-based relationship between A and B: they tend to happen close together in time
2. There is a time-based relationship between A and B such that B often happens shortly after A.
3. When I (or bot) creates condition A, then B usually happens.
4. When I (or bot) creates condition A, then B usually does not happen.
5. Science or simulations explain how A triggers B, if #2 is observed.
A bot can be programmed to conclude there is a potential causal relationship if #2 happens. If not problematic, the bot can then do experiments to see whether #3 or #4 is the case. If #3 happens, the bot can label the relationship as "likely causal". If #4, label it "probably not causal, but puzzling".
#5 would probably be needed to conclude "most likely causal", and is probably an unrealistic expectation for the first generation of "common sense" AI, although they may have a simple physics simulator built in. The highest "causal" score would be #2, #3, and #5 all true.
People getting their umbrellas ready is caused by the expectation of rain.
"Why are they taking their umbrellas out?" "Oh it must be because it is about to rain." "Why do they have an expectation of rain?" "Because they saw the weather report and clouds are visible"
For example, if you hunt to the east and find good game, you'd keep hunting to the east. Eventually you'd kill everything over there or get them to move, and your hunting would get worse. The optimal approach might be to randomize the direction you hunt, so that game doesn't learn where you're hunting.
The society couldn't say WHY the ceremony was good, but long term if they kept applying the ceremony they'd have better outcomes than societies that didn't.
Sometimes superstition is just that ... but I'd also bet there are unintuitive/surprising benefits behind a lot of it.
Obviously religions take advantage of this desire for the illusion of control and convince followers that practicing the religion will keep their lives free from external bad influences.
It had this context appropriate quote:
"One performs the rain sacrifice and it rains. Why? I say: there is no special reason why. It is the same as when one does not perform the rain sacrifice and it rains anyway. When the sun and moon suffer eclipse, one tries to save them. When Heaven sends drought, one performs the rain sacrifice. One performs divination and only then decides on important affairs. But this is not to be regarded as bringing one what one seeks, but rather is done to give things proper form. Thus, the gentleman regards this as proper form, but the common people regard it as connecting with spirits."
The more formal in the later process, the closer we reach to the real "knowledge". So your suggestion is quite practical, in the sense that we can start from that and figure out how to push the engine toward the more formal spectrum.
https://en.m.wikipedia.org/wiki/Structural_equation_modeling
Any human-like AI is still a "fake" - any notion of emotion, pain, empathy etc. we attribute to them is only a simulation. It simply doesn't matter. It amazes and amuses me to think that people might actually give a damn what the machine is "feeling". I think people who truly believe this are out of touch with reality and frankly, with other human beings. The machine doesn't really care about us, it's a bunch of ones and zeroes no matter how you slice and dice it.
Even after training them on cause and effect, they still don't care. I don't buy the "if it looks like a human, sounds like a human, it's human" argument at all.
Edit: alternatively, perhaps you only mean that any AI that realistically speaking might eventually be made would not be a person, not that no AI that could in theory be made would be?
Where do other mammals fit in your schema? Do you believe in some sort of biological essentialism? Or are only humans with caring about?
Machines deserve no such empathy, just a flick of the power switch as required. And I watched Automata (2014) last night. Didn't feel a single thing for the robots, as much as the film tries to goad you into doing so.
Edit: whoops, didn’t see the other reply
The preamble is depressing, since the episode aired right after a mass shooting, but Pearl gives a brief overview of his thinking.
any hooo....
Seriously. Extraordinary claims etc etc. If you want to claim humans are not solely material, you need to give some sort of evidence of a phenomena beyond the physical. You can't use intelligence per se as your evidence as then your argument is circular.
>Of course the Earth is in the center of the solar system. We must occupy a privileged space in this universe. Its been demonstrated for millennia.
Thinking is done with neurons. Neurons are subject to the same physical laws as the rest of the material world. Therefore thinking can be done by a machine (if nothing else, by a physics simulation of neurons).
To refute this logic, you must show some thought is not being done by neurons or that neurons are not subject to physics.
It might make for nice rhetoric, but it is more than a little disingenuous. Not only was that geocentrism piece not an argument or claim that I've made, but I've never seen anyone make it in that manner either. But perhaps you know that and were intentionally misrepresenting their arguments.
Back on topic, no, your concluding claim is not true. To reach your conclusion you would, at the minimum, need to assume that a machine can simulate arbitrary physical phenomenon, which is not a forgone conclusion. For instance, the "thinking done by neurons" you refer to made be reliant on some facet of the real numbers that is simply not computable. Perhaps at any level of approximation of it, what we discern as AGI may not manifest. Etc.
But, finally, your premises are faulty. What most people really mean when they reference AGI or "truly intelligent" is not intelligence, but wisdom. Computers have been "more intelligent" than humans for a long time now if it simply means arithmetic and recalling trivia. Now, noting that, isn't it possible that such wisdom is dependent on dependent on the will, the soul, etc?
So we arrive back at the true point of contention - you are a materialist. I claim that materialism has been handily refuted for thousands of years. Then you fell back on pretty much every freshman level fallacy in the book. On the other hand, I suspect some of your misrepresenting of "classical" philosophy was not intentional, and just the result of getting most of it second hand from the Kurzweil (AI) and Dawkins (geocentrism) type literature. It is wise not to be so dismissive of pre-"Enlightenment" thinking.