Slashdot Mirror
Opinion: Artificial Intelligence Hits the Barrier of Meaning (nytimes.com)
Machine learning algorithms don't yet understand things the way humans do -- with sometimes disastrous consequences. Melanie Mitchell, a professor of Computer Science at Portland State University, writes: As someone who has worked in A.I. for decades, I've witnessed the failure of similar predictions of imminent human-level A.I., and I'm certain these latest forecasts will fall short as well. The challenge of creating humanlike intelligence in machines remains greatly underestimated. Today's A.I. systems sorely lack the essence of human intelligence: understanding the situations we experience, being able to grasp their meaning. The mathematician and philosopher Gian-Carlo Rota famously asked, "I wonder whether or when A.I. will ever crash the barrier of meaning." To me, this is still the most important question.
The lack of humanlike understanding in machines is underscored by recent cracks that have appeared in the foundations of modern A.I. While today's programs are much more impressive than the systems we had 20 or 30 years ago, a series of research studies have shown that deep-learning systems can be unreliable in decidedly unhumanlike ways. I'll give a few examples. "The bareheaded man needed a hat" is transcribed by my phone's speech-recognition program as "The bear headed man needed a hat." Google Translate renders "I put the pig in the pen" into French as "Je mets le cochon dans le stylo" (mistranslating "pen" in the sense of a writing instrument). Programs that "read" documents and answer questions about them can easily be fooled into giving wrong answers when short, irrelevant snippets of text are appended to the document.
Similarly, programs that recognize faces and objects, lauded as a major triumph of deep learning, can fail dramatically when their input is modified even in modest ways by certain types of lighting, image filtering and other alterations that do not affect humans' recognition abilities in the slightest. One recent study showed that adding small amounts of "noise" to a face image can seriously harm the performance of state-of-the-art face-recognition programs. Another study, humorously called "The Elephant in the Room," showed that inserting a small image of an out-of-place object, such as an elephant, in the corner of a living-room image strangely caused deep-learning vision programs to suddenly misclassify other objects in the image.
The lack of humanlike understanding in machines is underscored by recent cracks that have appeared in the foundations of modern A.I. While today's programs are much more impressive than the systems we had 20 or 30 years ago, a series of research studies have shown that deep-learning systems can be unreliable in decidedly unhumanlike ways. I'll give a few examples. "The bareheaded man needed a hat" is transcribed by my phone's speech-recognition program as "The bear headed man needed a hat." Google Translate renders "I put the pig in the pen" into French as "Je mets le cochon dans le stylo" (mistranslating "pen" in the sense of a writing instrument). Programs that "read" documents and answer questions about them can easily be fooled into giving wrong answers when short, irrelevant snippets of text are appended to the document.
Similarly, programs that recognize faces and objects, lauded as a major triumph of deep learning, can fail dramatically when their input is modified even in modest ways by certain types of lighting, image filtering and other alterations that do not affect humans' recognition abilities in the slightest. One recent study showed that adding small amounts of "noise" to a face image can seriously harm the performance of state-of-the-art face-recognition programs. Another study, humorously called "The Elephant in the Room," showed that inserting a small image of an out-of-place object, such as an elephant, in the corner of a living-room image strangely caused deep-learning vision programs to suddenly misclassify other objects in the image.
I wonder if these AI vision systems that input millions of images are actually doing a deep learning, or are just canvassing pretty much every image possibility such that any possible live image is just a tiny automated delta calculation away from an answer.
This would explain why tweaking the input in the described ways would throw the AI into a tizzy -- the tweaked input isn't within a tiny delta of any of the millions of categorized images.
(-1: Post disagrees with my already-settled worldview) is not a valid mod option.
"...programs that recognize faces and objects, lauded as a major triumph of deep learning, can fail dramatically when their input is modified even in modest ways by certain types of lighting, image filtering and other alterations that do not affect humans' recognition abilities in the slightest." Now tell me again what a great idea self-driving cars are!
I've abandoned my search for truth; now I'm just looking for some useful delusions.
As human beings, we like to _pretend_ are decisions are guided by pure reason, but in actual practice we are driven more by tribal instinct, in exactly the same way that our dogs are still driven by pack instinct. Man as a completely rational being is a myth. Perhaps not being able to relate to irrational instinctual thought is the real barrier between men and machines understanding each other, rather than the machines lack of a physical body. Machines can now emulate all human sensory input, given enough time they should be able to develop a similar model to humans of what operating in the physical world "feels" like.
I've abandoned my search for truth; now I'm just looking for some useful delusions.
In the example given all that is needed here is better pattern recognition which is really what we associate as meaning. If you say "pen" in a sentence referring to a pig, sheep etc. then we naturally tend to assume pen=small field. There is no reason that an AI cannot learn that through better pattern recognition i.e. more training with better algorithms. The AI can certainly know that 'pen' refers to different possible objects, just like we do, but if you talk about animals then our pattern recognition triggers the "small field" meaning and if you are talking about writing then it triggers the "ink-related" meaning.
Of course, it will need really good training and algorithms to figure out sentences like "I wrote about the pigs using my pen." but there is no reason to assume that there is some barrier to AI doing that. The compsci department round the corner has colleagues working on text and speech recognition and I'm sure this type of thing is something they are dealing with and I doubt Google translate is that close to state-of-the-art.
Gross oversimplification.
“Common sense is not so common.” — Voltaire
But there is no denying we love boobies.
About 140 billion neurons in human brain. Grossly oversimplifying 140 billion ^2 possible interconnects (actual number is lower). We can't even store state information for the synapses (input weights), much less model the chemistry in the synapse.
John McAfee 'It was like that time I hired that Bangkok prostitute; to do my taxes, while I fucked my accountant'
This is all old school and nothing new. Computers advanced to the point where people realized they could practically use it. Neural networks are what brains use. Biological brains though have networks of networks. Neural networks are like fourier transforms. They identify a signal from noise. They work on corelations though and set data. They are literally educated guessing machines.
A real brain has neural networks that work together in sets. And on top of that there is a genetic cheat sheet for the neural nets; how big they are and how they should feed back into each other. There are even neural nets active in youth that function as trainers or biasing to boot strap brains. An insect has more intelligence than modern implementations. Modern systems are more akin to the pre and post processing that occurs locally in the optic nerve and spinal cord.
The big snake in the grass is the term Intelligence. It is a fuzzy concept in itself that depends on context.
The Turing test has led us down a rocky road and we have a very long way to go. Artificial human-like intelligence IMHO is still a long way away. Most people make shoot from the hip assumptions about how the brain works and after doing some basic math about Moore's law assume super intelligence is right around the corner.
The brain is way more complicated than we know.
For example: there are two stable isotopes of lithium. Chemically they are identical, but they do not have the same effect on the brain. One is useful as a drug to treat mental illness and the other is not. This means there is something more subtle about how our brain works than interconnections and electrochemistry.
It is however a worthy challenge because the journey will teach us much about who we really are and how we work.
Greed is the root of all evil.
Humans also don't get trained on still pictures. Babies recognize objects because they move in front of other objects. Babies learn early that objects, that are temporarily behind other objects, tend to come out again, if the movement continues. You can test that with the eye movement of babies, which follows the imagined path of the moving object until it arrives in sight again after it has passed the hiding object. Babies also notice early, that their own movement causes objects to change position in their visual field, and that each object has an unique way to change its position during movement, depending on the point of view of the baby relative to the object. Thus long before babies are able to identify objects as chairs, tables or toys, they are able to tell objects as such apart, because each object moves differently in their vision when they move the head.
You don't get this training by showing still pictures to a computer. You should a) use movies, and b) give the computer the ability to move its point of view within scenes to learn how to tell objects apart. But that's much more complicated than having the computer process vast stacks of annotated pictures.
By not passing protectionist tariffs that crippled half the states into law.
An enigma, wrapped in a riddle, shrouded in bacon and cheese
Thanks for the link. I read the article and many of the comments.
What do you think about this one?
The same thing I think about anyone who claims that a major moment in human history boils down to 1 factor - it's bullshit, man.
Yes, slavery was a factor, but not the only factor. Consider the tariffs I linked to, then ask yourself: under those trade rules, how would the Southern states have managed to survive without the use of slave labor? The fact is, they wouldn't have, so in a way the Northern states forced the South to rely on slavery, then punished those states for it.
The fact is, our American Civil War was complicated, both the reasons for it's beginning and continuation (fun fact - Lincoln floated the idea of leaving slavery legal in some states, to preserve the union). What's interesting as an American is that the angle historians take on the conflict tends to be defined by where you get the education: Northern states tend to teach the "Civil War was about slavery" concept, Southern states lean towards the "state's rights" ideology, and Border states (like where I'm from) tend to take a more middle-of-the-road, "both of you are assholes" mentality.
An enigma, wrapped in a riddle, shrouded in bacon and cheese
Entropy. Compression. Same thing. The whole world is thermodynamics and your state of knowledge about the world is also limited by thermodynamics. There will never be a computer that can predict the future of the universe before the universe arrives simply because you can't store a representation of the universe inside the universe itself.
Lossy Compression is therefore how we get around that and be able to compute/think/predict what an approximate future state of the universe is.
What the goal is to align the losses of the compression into the input space which does not exist. For example, if there is no possibly image of a living room of size smaller than an elephant that could contain the elephant then any mapping of images with and without elephants to the same compressed reprensention is a good compression. To say it differently the compresses state is a many to one mapping back to the original state. If for every compressed state there is only one realizable original state then it's invertable. THe images in the original space that could never happen are also mapped to the same compressed state but because they could never exist we lose nothing by ignoring them.
Thus compression and prediction are the same thing.
AI fails when it either over-compresses to a space too small to hold every realizable state. Or it compresses poorly so that in unnessarily conflates two possible real states. For example, the uber car that thought the woman in the road was blowing trash.
On the otherhand, it's often very valuable to overcompress as long as you are tolerant of mistakes on the prediction. That is, the uber car in question was able to do a great job of driving most of the time because it made fact choices that were nearly always good enough. The Cheetah can't just chase the antelope, it needs to try to guess and cut corners a bit. As long as most of it's guesses are good it wins. In the case of the cheetah, a mistake just means a missed meal, which is tolerable. But in the case of the uber driver or an ICBM nuclear missile failsafe system, then our tolerance for error is a bit lower.
Thus a little overcompression is acutally good for generalizing rather than parroting.
A lot of overcompression leads to bad predictions.
Some drink at the fountain of knowledge. Others just gargle.