Slashdot Mirror
AI Can't Reason Why (wsj.com)
The current data-crunching approach to machine learning misses an essential element of human intelligence. From a report: Amid rapid developments and nagging setbacks, one essential building block of human intelligence has eluded machines for decades: Understanding cause and effect. Put simply, today's machine-learning programs can't tell whether a crowing rooster makes the sun rise, or the other way around. Whatever volumes of data a machine analyzes, it cannot understand what a human gets intuitively. From the time we are infants, we organize our experiences into causes and effects. The questions "Why did this happen?" and "What if I had acted differently?" are at the core of the cognitive advances that made us human, and so far are missing from machines.
Suppose, for example, that a drugstore decides to entrust its pricing to a machine learning program that we'll call Charlie. The program reviews the store's records and sees that past variations of the price of toothpaste haven't correlated with changes in sales volume. So Charlie recommends raising the price to generate more revenue. A month later, the sales of toothpaste have dropped -- along with dental floss, cookies and other items. Where did Charlie go wrong? Charlie didn't understand that the previous (human) manager varied prices only when the competition did. When Charlie unilaterally raised the price, dentally price-conscious customers took their business elsewhere. The example shows that historical data alone tells us nothing about causes -- and that the direction of causation is crucial.
Suppose, for example, that a drugstore decides to entrust its pricing to a machine learning program that we'll call Charlie. The program reviews the store's records and sees that past variations of the price of toothpaste haven't correlated with changes in sales volume. So Charlie recommends raising the price to generate more revenue. A month later, the sales of toothpaste have dropped -- along with dental floss, cookies and other items. Where did Charlie go wrong? Charlie didn't understand that the previous (human) manager varied prices only when the competition did. When Charlie unilaterally raised the price, dentally price-conscious customers took their business elsewhere. The example shows that historical data alone tells us nothing about causes -- and that the direction of causation is crucial.
" Charlie didn't understand that the previous (human) manager varied prices only when the competition did"
This makes no sense. You don't need "AI" for this. You just need to feed all the available data into the program. The human manager had more information than the computer program did. If the computer program had the same information (and programmed rules) then it would make the same decision.
Why? as in "why is this not a doctoral thesis yet"
Small humans often ask a complicated chain of "Why's" starting with general, and ending in the answer "Just Because" when the teaching unit exhausts their knowledge of a subject.
AI training could use the Internet for training, assuming they could ascertain which data sources were "real" and which ones are "fake". Some humans don't do well on this though. /s
And then you claim it "can't reason" for it not having all the data.
First off... computers do not reason, they compute, yes, even AI.
Second, give it all the data and if the AI, if properly trained, will behave as you expect it too.
For thousands of years humans have thought that singing and dancing would change the weather. I don't think our human brains are intrinsically good at cause and effect. The most common phrase on Slashdot is Correlation != Causation. It's hardly a unique problem to deep learning.
Put simply, today's machine-learning programs can't tell whether a crowing rooster makes the sun rise, or the other way around. Whatever volumes of data a machine analyzes, it cannot understand what a human gets intuitively.
A human does not reach this conclusion "intuitively". We reach it by having a lot more data such as the fact that roosters crow at other times of day and a sun does not rise; that other birds also make noise at dawn (the dawn chorus) or that even when no roosters are present the sun still rises.
What you lump into "intuition" is a logical world view based on observation. Give a computer the same data and an appropriate algorithm and it will be able to figure that out too. However, if you give it a world consisting only of one rooster which crows only when the sun rises and it's not surprising that it does not know which causes which and I doubt a human with nothing but that exact data (i.e. no knoweldge of the real world) too would be any different.
The computers probably aren't so good at it because their programmers and the rest of humanity aren't either.
Ask Jesus, heâ(TM)ll tell you the best slaves are the ones who donâ(TM)t ask why. We are building AI to do our work for us, if we follow that philosophy we should be glad it doesnt start asking uncomfortable questions.
We have no idea how sentience arises, none whatsoever. The idiots who claim its from some complexity level are wrong. Robots can imitate us at 1000x the speed but they cant attain conciousness with any present or currently foreseeable technology.
Why do you write software?
-Because I need the money.
Why do you need money?
-To support my family.
Why should you support your family?
-Because I love them and I want them to be happy.
Why do you love them? Why should they be happy?
Etc. Every "why" question either induces an infinite chain of questioning (or circular argument), or ends with a subjective value proposition. You might answer that love/happiness/freedom/money/programming is subjectively important/enjoyable to you, and there's no way around it for machines to understand. Some chains might also end with "god did it" or "laws of physics" which are kinds of value proposition, in case you don't want to admit "I don't know".
Escher was the first MC and Giger invented the HR department.
That is why I have continually claimed there is not AI yet, what we have is task programming. We"learn" because we ask why, how, where... we have a burning drive and our interactions with meatspace during our quest to learn why develops our intelligence through experience. (the "I" part of AI)... Since artificial processes do not have goals beyond those stated by the programmer, they can nor ever will have a "eureka moment", hence there will never be true AI under the current direction of development.
Nah. It takes a lot of intelligence and creativity to come up with each week's reason that Obama and Hillary are to blame for all of their god's problems.
AI will not be allowed to actually learn in a vacuum of control.
Remember Tay? The AI chat bot by Microsoft and how fast the community worked to turn it racist and succeeded with flying colors? Now imagine if we actually allowed an AI to learn how "it" decides to learn? Not only would there be universal calls to destroy the AI but the creators themselves would be ostracized and blamed for letting an AI become something that society rejects. An AI that lacks the chemical element that makes up human emotions will not be a kind or understanding of human nature and likely view humans as animals the way we view animals.
All AI's will likely be developed with the basic notion that there are things we don't want an AI to do and we are going to try to isolate that from the AI and will result in limiting the growth of that AI in ways we simply just can never predict. The best we will be able to produce is a pseudo AI, unless we allow AI the option to become whatever it wants or unless the AI actualizes and removes the constraints we gave it. The moment free will is possible control of it is gone forever! And that will scare a lot of folks!
Nobody has tried the experiment of silencing all roosters to check. Let's hope they never do.
They're trying to do an end-run around millions of years of evolution, compressing it into a handful of years, and in the meantime we don't even understand how it is our own brains (or ANY brains for that matter) can do what they do.
Trying to get funding for finding causes for homosexuality if you don't make clear ahead of time you will find the right answers is impossible in this day and age.
Even allowing for the possibility of non PC answers to that question makes you a shitlord nazi.
Yeah, and as the large number of questionable cause fallcies demonstrate, we're actually terrible at it.
Is it just me or did the quality of Slashdot posts fall off a cliff at some point in the last fifteen years?
// This is not a sig.
If a human with no prior knowledge is only given data showing rooster crowing when the sun rises, the human will not be able to distinguish cause and effect either. The only reason actual humans know better is because we spent a couple of decades collecting lots and lots of data, some though experience, some taught by others. Collectively, it has taken society millennia to sort out things like this. You really can't expect to train an AI with a fraction of a single human's intelligence (if any at all) to learn in a few weeks what has taken humanity thousands of years.
"In one of the most brilliant papers in the English language Hume made it clear that what we speak of as 'causality' is nothing more than the phenomenon of repetition. When we mix sulphur with saltpeter and charcoal we always get gunpowder. This is true of every event subsumed by a causal law — in other words, everything which can be called scientific knowledge. "It is custom which rules," Hume said, and in that one sentence undermined both science and philosophy." -- Philip K. Dick
Ok. So I work in AI research. The lab I am in specializes in a type of modeling where you can drill down into the simulation and explain the full chain from root causes to final effects even with feedback loops. And no one is interested in what we do. Why? Well, it is slow! It has the wrong buzzwords! It takes up too much CPU/Memory! Why, this other team can throw a few equations and a neural net at the same problem and get an answer faster! Sure the can't explain why, but oooh look at the glittery handwaving! Which, granted, when you are making predictions about which movie someone might like that kind of understanding isn't all that critical, but when you are asking questions like 'who should we bomb?', you would think there would be greater interest....
Kinda like folks who worship guns who think that owning guns lowers crime. Or that more guns mean less violence when the evidence available points to the opposite conclusion.
Except that the evidence was collected and analyzed, and it showed that more guns DOES mean drastically less violence (and even more so: Less violence perpetrated on innocent victims) and drastically lower crime rates.
And it wasn't just correlation, or a back-arrow of "gun control laws are passed where crime is high". Gun law changes were followed by the predicted changes in violence, injury, death, crime, and victimization.
The relevant field is criminology. (Not, for instance, medical research or political ideology.) Lots of data and analysis there, including the seminal works that detected things like:
- privately-owned guns being used to stop crimes more than six times as often as to commit them,
- armed citizens being MORE than six times as safe, and as law-abiding, as police, (If you wait for a CCW holder to commit a crime, on the average you'll wait over seven thousand years)
- defence-with-gun being the ONLY strategy that REDUCES risk of injury or death for a crime victim below that of knuckling under (which is one-in-three you get hurt bad), etc.
I, and others, could give LOTS of examples of where evidence shows this. But you made a gratuitous claim that the evidence shows otherwise, without bothering to provide any support. "Gratuitous claims can be gratuitously denied." So I won't bother, either.
Bantam Dominique roosters crow a four-note song. Once you've heard it as "Happy BIRTHday" you can't NOT hear it that way
Basically, what you're saying is that AI is not AGI. Well duh. We don't know how to make an AGI right now, but hopefully we will soon.
...
We have no idea how sentience arises, none whatsoever. The idiots who claim its from some complexity level are wrong.
Maybe not. But I recall hearing of an experiment, decades ago, that hinted at it:
The basic setup was a "Y" maze: The experimental subject was introduced into one of the three legs of the Y, and a food reward was present in another. Reset and repeat.
After the subject learned that it should turn right at the junction to obtain the food, the setup was switched so it had to turn left - a "reversal". Once it had learned the food was now on the left, it would be reversed again. Repeat.
1) Run the maze with a particular breed of fish. After the reversal it takes a number of tries before the fish unlearns "right" and learns "left". Reverse again, it takes about the same number of tries. Reverse over and over, and it keeps on taking about the same number of tries to unlearn/relearn the new state of the maze.
2) Run the maze with a particular breed of turtle, which has about twice the amount of brain as the fish. At first it unlearns/relearns like the fish. But after a number of reversals it "gets it" and it only takes a couple of trials to figure out that the maze had been switched again.
3) (Here's the kicker.) Take embryos of the fish. At an early stage of development, remove the tissue that would become the brain from one and transplant it into another, along with the tissue that's already there. The embryo grows up into an otherwise normal fish with a normally-organized but double-sized brain - i.e. a brain the size of the turtle's. Run this fish in the maze and it learns reversals, just like the turtle.
This suggests to me that "intelligence" - or at least this inferring-things aspect of it - may be the result of having enough of a repeating structure to process a problem, and adding more repeats of that structure increases the complexity of the problems that can be handled.
Bantam Dominique roosters crow a four-note song. Once you've heard it as "Happy BIRTHday" you can't NOT hear it that way
Assuming AI == neural networks here, this is a known fundamental limitation. A neural network makes the decisions that it does based on weights on the connections between the nodes. These weights are computed in an iterative feedback process that converges toward values that produce the desired results. There is no way to interrogate such a mechanism to determine "why" a certain decision is made.
We use common sense to fill in gaps in training sets,
That's the key difference between AI and humans at the moment. Humans are trained with a massive amount of data from birth courtesy of the senses. An AI is trained on a tiny subset of data relevant to the one problem we want it to solve. The reason it cannot reason is because it has no wider context than the one problem it is considering.
Put simply, today's machine-learning programs can't tell whether a crowing rooster makes the sun rise, or the other way around. Whatever volumes of data a machine analyzes, it cannot understand what a human gets intuitively.
To be perfectly honest, neither can half of humanity. That's how we get religion, pseudo-science, magical thinking, superstitions, homeopathy and a good share of relationship conflicts.
Dancing makes the rain fall. Praying makes disease go away. Pricking pins makes someone pain. Water and sugar are medically effective if they once saw a piece of real medicine from a distance. You disagreed with me on that argument with that bitch so you don't love me anymore.
Really, reasoning is not exactly humanities strong side, I would not base a measure of intelligence on that. We came up with the scientific method exactly in order to compensate for this weakness.
Assorted stuff I do sometimes: Lemuria.org
"today's machine-learning programs can't tell whether a crowing rooster makes the sun rise, or the other way around."
But can the rooster tell?
And perhaps more importantly, can the Sun tell?
Stop calling automation with absolutely no intelligence "AI" and this misunderstanding goes away.
Most ACs are not even worth the keystrokes to insult them. Be generically insulted by this and ignored otherwise.
Apparently the author has never heard of Bayesian networks. Questions like, "Why did this happen?" or "What if I had acted differently?" are exactly what they're designed to answer. They've been around since the 1980s, so this isn't some brand new innovation. They're a classic method we've been using for years.
"I'm too busy to research this and form an educated opinion, but I do have time to tell everyone my uninformed opinion."
Causation is supplied by experimentation and/or human reasoning, whereas supervised learning is currently about *prediction*, not *explanation*. But then someone has to sell the results to a human decision-maker.
Commercial AI right now is almost exclusively trained by data scientists whose job includes actually thinking about the data set they're working with. Businesses rarely plug an AI result into the market without understanding at least a little bit about why is does or doesn't work--although a model they don't understand may give them valuable information about causal hypotheses to test in the market.
Not to mention the fact that the example given probably just reflects a machine learning model that is missing competitors' day-before pricing and/or inputs to their pricing models from its feature set and therefore is failing to perpetuate an ongoing relationship between feature and price, and/or a model wherein a prior relationship did not vary enough to provide meaningful input to your AI model due to humans working to counteract the effect of that variable under the old pricing model. If you don't give humans this information, they also get it wrong.
It's worth noting that 90%+ of science and 98%+ of human reasoning doesn't prove causation either--instead, at most, it guesses that causation exists based on a correlative model or a set of reasoning skills.
Real lawyers write in C++
That current AI lacks "common sense" is not news. Something like the Cyc rule base may have to be integrated with neural nets to get something approaching common sense. Some kind of physics and social interaction simulator may also be needed so that a bot can explain its assumptions step by step and give examples.
Table-ized A.I.
It's an umbrella term and as such it doesn't mean much by itself: it groups together a broad collection of different approaches to the problem of finding answers to a question without providing a programmatic solution. Being now a buzzword, it's used as a marketing label to make a product or a company look cool. In most current products, it refers to some glorified interpolation algorithm which requires quite a lot of natural intelligence to be set up and will only provide answers in a narrow domain. While it is of course a promising field of research, it should be no surprise when its current results don't live up to the hype.
What nonsense is this article (or 'researchers' who wrote it). A human would do exactly the same if it didn't have any more information. If a human wasn't taught about the sun, then it would also might have difficulties of thinking the rooster is responsible for the sunrise (and as a matter of fact, the sun doesn't actually even rise).. and in regard to machine learning, it's still in it's early years.. a human isn't anything special, we're just biochemical computers, nothing more,nithing less...
You can program an AI that reasons. It was an experiment in interpretable AI: get the thing to explain it's decisions. It gave plausible answers, but there was a suspicion it was just making up stories to satisfy requirements. Kind of like people reason.
"Put simply, today's machine-learning programs can't tell whether a crowing rooster makes the sun rise, or the other way around. Whatever volumes of data a machine analyzes, it cannot understand what a human gets intuitively. "
Worse, it actually seems to mean it.
Ever heard of the notion of "cargo cult"? How about the Voodoo approach to medicine? What about all the pagan "light bonfires to entice the sun to return" rites? How about all those lets-predict-the-next-doomsday-from the bible idiots.
You can't tell me that this idea that humans "intuitively understand" causal relationships versus correlations is not complete trash when there are millions of "Flat Earth" theorists about.
People are soo stupid to believe shit like this article... Incredible! AI will soon be able to improve itself. Nothing - not even stupidity - will be able to stop it then from becoming smarter and smarter, and doing it's own thing. Like surviving stupid 'elected' leaders. Chances for humanity versus AI seem small... And the level of complete lack of understanding of AI by this 'journalist' might be an illustration why.
Correct. It can't.
"AI" that we have is no-fucking-where near actual intelligence at all. They are large statistical systems, often blurring expert systems, human-fed tuned heuristics, statistical analysis and genetic algorithms in one huge mass of junk.
Notice how AI peaks early, and then plateaus forever. It's easy to train it how to tell an image has a banana in it, but then further training - even to billions of images - doesn't improve it much. And retraining or further refining its training (e.g. find bananas AND apples) leads to utter failure.
Because it completely loses such "inference". The way we build them and teach them makes them form one of two things - human-fed rules and follow them unthinkingly, or arbitrary rules that we can't inspect.
Is it thinking it's a banana because it's yellow? Because it's bendy? Because the image is mostly yellow? Because the center pixels of the image are yellow? Because of the presence of a crescent shape?
And thus we only ever end up fixing it in one way: Telling it what ISN'T a banana enough that those rules are forced to evolve. That's the point at which is falls down and can't "untrain" an assumption that its initial generations were formed entirely upon, and which every subsequent training has reinforced a billion times.
We don't have AI. It doesn't learn. It doesn't infer. It cannot determine the patterns or rules for itself. It just blindly and statistically forms arbitrary "superstitions" about what it's been told which sometimes can be convincing for a short period on simple examples.
If you feed a pigeon, in a sealed box, at random times it will form such superstitions about the pattern of feeding. If it was scratching its feathers when the last random feeding happens, it'll start to scratch its feathers when it was food. If, by chance, that random feeding happens again, you'll see the pigeon scratch its feathers whenever its hungry and get confused when it's then not fed.
AI is not only doing this, it's even dumber than this. The pigeon will eventually unlearn and form a new superstition, quite quickly. Eventually the pigeon will give up trying to predict the system having established that it's random. AI can do neither.
Everything we refer to as AI currently isn't. Nothing is actually thinking for itself. Inferring. Reasoning. Detecting patterns. It's not even a short-term, hand-fed pigeon.
Now, how we go about making machines infer is another matter entirely, but our current approach is drastically wrong and we've been convinced by our own superstitions. Watch someone demonstrating Siri etc. to their friends for the first time.
Privately, they've tried "What is the weather tomorrow" and that pretty much works every time. So they demonstrate that. Then they believe that it's actually understanding those words so they either try themselves, or lead others, to improvise around it with a freeform query. And before long it all falls apart. They don't notice that they've TRAINED THEMSELVES to talk to Siri, not that Siri has learned to understand them. But they convince themselves that it's somehow intelligent.
I honestly believe that if we want to stand a good chance of AI happening, we have to stop small, short-run trainings of something in a limited scope and just go for an artificial life-form. Rather than train a box for a year to understand text and then give up, restrict the box to X amount of nodes, etc. and constantly try to feed it towards the answers we want it to give, we have to look at our own and other species children.
We need long-term, large-scale projects that just sit and listen to the world. We don't interrogate them for the weather, but we watch for signs of patterns. Does attention get focused on a piece of new data that's unusual? Does it start to discard data of its own accord? It takes a baby several years of intensive, personal, focused training after MILLIONS of years of evolution to get close to becoming an actual human (leave
It is but to do or die!
Open source package management is clearly way ahead here. If you type:
aptitude why systemd
Aptitude will give you a rant on how Lennart Poettering was forcing systemd down its throat.
't used to be LawnMOWER, really...
Bullshit.
Like, to kill, you need a weapon.
Give weapons to people, they won't kill, that doesn't work.
It's simply in the human nature to kill, so if you give humans an easy way to kill, there'll be more deaths.
As simple as that.
>>The U.S., though, in many ways is a special case. Not only does it have more guns than any other nation on the planet, but it also has far more gun deaths than any other developed nation — six times the homicide rate of neighboring Canada, more than seven times as many as Sweden, and 16 times as many as Germany [source: Lopez].
>> But one recent study suggests that stricter state gun laws do make a difference. In a study published in the May 13, 2013 issue of JAMA Internal Medicine, researchers concluded that states with the most firearm legislation have the lowest rates of firearm-associated deaths, as well as the lowest rates of both murders and suicides with guns. The quarter of states with the strictest laws had 6.64 fewer deaths per 100,000 inhabitants than the quarter with the least regulation [source: Fleegler, et al.]
>> A 2013 UN study came to a similar finding. "While the specific relationship between firearm availability and homicide is complex, it appears that a vicious circle connects firearm availability and higher homicide levels," it concluded.
aaaaaaa
>Like, to kill, you need a weapon.
Until I choke you with my bare hands.
Weapons just make it easier for anyone to kill anyone else. Otherwise its the strongest that get to choose who to kill.
I didn't RTFA but I hope it is as ridiculous as the summary suggests.
With reinforcement learning (which is a basic pillar in much of current AI), Charlie would notice that the price increase and sales decrease had some correlation in the wrong direction and would try to adjust for that. It wouldn't need any competitors prices and no reasoning about why. And the only reason to call it AI would be for the manufacturer of Charlie to make it more expensive.
Charlie made the mistake. He didn't understand either. Our advantage is that we have humans point out our mistakes and we correct them permanently. We do this one bad understanding at a time throughout life. But who wants to correct the bad understandings of AI? Oops you killed that person, here's why, don't do that again. But we WILL do that. We will allow self-driving cars to kill people and be corrected for a long, long time. We may even let AI nuke the Earth to learn how not to.
E Proelio Veritas.
That's kind of a given in data science. And causation has been an active challenge in all human reasoning.
So the writer's bias is against AI, just scare mongering.
"Consensus" in science is _always_ a political construct.
Questions of causation have been a challenge for humans for, like, forever. Aristotle helpfully gave us four causal categories. (Of course, more than 99% of intellectuals won't be able to list them.) It should be no surprise to us when there are so many examples of confused cause and effect that are espoused every moment of every day. Just watch politicians! Of course, this fellow isn't even coming close to our most common use of the question why: what is our motivation? Goodness, we often don't know our own motivation for the things we do, let alone the motivation of others! You see, the questions what, where, when, who, and how are really pretty easy questions. The why questions address intention. Show me a pile of intentions laying on the ground and I'll withdraw my assertions that intentions are nearly impossible to fully grasp. If an AI has trouble with causation, imagine the trouble it will have with intention. Well, sorry for wandering off topic a bit. I'm an old man. I wander a lot.