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.
"Orange man bad", but they know not what it means.
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.
I recommend the book "philosophy in the flesh" as a starting point.
The upshot is: the concept of "pure reason" is a fairy tale. There is no such thing. All human reason is rooted in our most basic experience of being "embodied." Meaning we have foundational concepts such as movement-towards, distance-between, contained-within, breaking-through, etc., that are a direct result of being things with physical bodies, and that serve as an inescapable cognitive foundation for all our higher reasoning. In fact, basically all of our lofty and abstract thoughts are metaphors back to these base concepts.
Any cognitive engine that attempts to be all reason without these base concepts will fail to understand. It won't think as humans do.
This can be overcome....if a brain can do it then so can a network of transistors. But the process will require some kind of instilling of these fleshy base concepts in order to get the results that we expect.
"...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.
generalization over situations, and bayesian statistics?
I think the issue is that the AIs have not experienced / perceived / taken in data about enough different kinds of situations, and specifically, have not been aimed at the problem of "what if I am an agent with goals in all these different situation types."
Right now in AI, mostly we are training the "visual cortex" or the "language parsing centre" of the brain.
The algorithms are not being applied to the general agent problem. The low hanging fruit of constrained commercializable sub problems is being covered first.
Where are we going and why are we in a handbasket?
There's a DARPA BAA on the street right now for "Machine Common Sense" that is hoping to address this by asking AI researchers to design AI to learn "common sense" the same way human babies do. One of the examples in the text of the BAA is "I saw the Grand Canyon flying to New York." A context-aware AI or one with "common sense" would understand that this sentence really meant "...WHILE flying to New York" rather than inferring that the Grand Canyon was flying.
"Common sense" in DARPA's context is not really what I would call the widespread understanding of what that phrase means, but is more oriented around understanding basic physics and behaviors and recognizing when something doesn't make sense. They're also taking...um....baby steps with this BAA, just trying to get some basic behavior around recognizing un-physical scenarios and that sort of thing. It's pretty cool though.
Read about the BAA here. . Download the ~1.9MB PDF for the full text of the BAA.
Artificially intelligent systems don't have intelligence, machine learning systems don't learn anything. Why? Because intelligence and learning rely on understanding, and the algorithms don't actually understand anything, they are just some lines of code.
Now the folks on the AI gravy train may try and modify the definitions of intelligence and learning to make it look like they are achieving their goals, but we know they are not.
First psychologists/cognitive linguists thought the brain worked like a machine, then a circuit, then a computer program. Each time, they find out that their rules cannot describe certain human behaviors.
Now we're finding that linear algebra-based models can't handle nonlinear sequences (i.e. "out-of-context information"). A human can categorize the extra information as necessary or not right away, but computers have to process it first to figure that out. A logic bomb for AI! Nice.
That's where we are at. People have to realize that nowadays we actually have Fake Intelligence, and as any "fake anything" has a ton of undesirable side-effects and shortcomings. Deep learning and similar techniques won't advance real machine understanding. They are simply shortcuts/cheats to achieve Fake Intelligence.
Of course, calling today's A.I. what it really is (Fake Intelligence) isn't going to attract many investors... so people keep calling it A.I.
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.
That sounds like progress from the days where the bleeding edge software couldn't even parse natural language. We essentially have the Babelfish at ~70% accuracy.
Everyone knows you can train a dog to be obedient (like sit or stay), do tasks (such as herd farm animals or sniff out specific objects), or do tricks (bark, run a course) etc.
Most people think cats are untrainable because you can't teach a cat to do any of those things, but cats are trainable, just with different things, because they're fundamentally different animals. They can be taught to fetch things, jump through hoops, and other things that play to cats' strengths as solitary hunters, whereas dogs can be trained to do more social things because they are inherently pack animals.
Maybe it's not correct for people working on AI to try and make AI do human-like things. Machines are different than humans, so why try to make them humans? AI should focus on optimizing what machines are good at, such as pattern recognition across vast data sets, and not trying to force them to do what they're not good at, such as contextual analysis.
This analysis is common knowledge for those who work in machine learning and AI. But given all the money flowing into businesses built around these technologies, the whole topic, and in particular words and phrases like "deep learning" and "cognition" get heavily overloaded / misused to exaggerate what is possible and to confuse discussion around those that are unlikely / difficult / impossible.
Yes you can train ML machines to do amazing pattern recognition. But it is still just pattern recognition: there is no cognition or understanding. None at all.
To know a technology well you really need to know what it is AND what it isn't. Otherwise it's easy to be fooled.
AI is sort of loosely used these days, it sound really good to use it but AI can be defined in many ways some of which are not very intelligent.
Take self driving vehicles, sounds great on paper a car driving you around, but even today that car cannot drive you around in bad weather. No fog, rain, snow etc. A lot of wishful thinking about AI these days incorporating into our daily lives. But how much is so over promised and over hyped it really isn't that intelligent?
But the tech-fanatics want their flying cars...
I also should add that there is no indicator at all that machines will ever get there.
Most ACs are not even worth the keystrokes to insult them. Be generically insulted by this and ignored otherwise.
Mixed: I would not be able to come to my own conclusion otherwise.
I believe at least two things will have to happen. First, the bot will have to generate candidate models of reality and evaluate them against the input for the most viable fit. These models may be physical in some cases, such as a 3D reconstruction of a face or room; conceptual in others, such as social relationship diagrams; and logic/deduction models, perhaps using CYC-like rule bases.
Second, these models and the rules that generated them will need to be comprehensible by 4-year-degree analysts so enough staff can tune the generation rules and/or the model templates. Obscure esoteric behind-the-scenes computations won't do. One has to know why a bot got a wrong answer to adjust it. Self-learning only goes so far; it's why our parents had to spank us (or a long time-out) when we ran out in the street. If they didn't, we'd be too dead to learn. Bots will probably need human teachers of some sort.
These techniques somewhat already exist in the AI field. But they need to be improved and integrated with each of the other AI methods. Dare I use the word "synergy"? Humans use many "clues" to triangulate interpretation and decision; some of them pattern-based experience, some of them closer to logical deduction. AI has a lot of techniques available, it's just that nobody has figured out how to glue them together and analyze them to easily tune and debug the "glue job".
Table-ized A.I.
I've been saying this for a while (just to my friends) - I think algorithms make these mistakes because there are no consequences for wrong answers. It is in the best interest of humans and living creatures to guess correctly, because we don't want to die. For example, tigers have "meaning" to us because they could kill us, food has "meaning" to us because we die without it, etc. etc. Nothing has "meaning" to predictive algorithms, which I think is a interesting and fundamental challenge to predictive modeling and machine learning in general.
The asphalt lobbyists don't want us to have flying cars. "Roads? Where we're going, we don't need roads."
But more seriously, there is a real desire for fast travel that isn't limited by long waits like in a train schedule or unpredictable travel time like in heavy traffic. If you take the whole "flying car" thing from mid-20th century Popular Science magazines overly literal, we're probably very far away from that. But we are moving towards technology that addresses similar demands for convenience and will have some of the same advantages of the theoretical flying car.
“Common sense is not so common.” — Voltaire
All our efforts to build machines of any type have always follow the same rules: accumulation of simple parts performing simple actions. And this doesn't just refer to the way in which machines work, but also to the whole process required to firstly build all of them. Computers, for example, didn't appear suddenly as a result of someone's happy idea, but were the result of centuries of learning of different aspects starting from knowing how to safely manage electricity. This has always been the case and I don't think that many people have ever doubted that reality before: step-by-step, by solving initially unrelated problems and by gradually merging isolated solutions into more comprehensive ones. The abstract idea of a computer as a machine performing calculations was probably easy to understand hundreds of years ago; but the current computers and all what is required to make them run was completely unimaginable. I don't think that people being amazed with the first computers counting up to 1 million in a few minutes were expecting them to eventually move to their current speeds. Or, at least, nobody back then in their right mind should have thought about that eventually being just "one magic leap" away.
Let's make it simple and think about what is required to create a machine able to somehow emulate human memory. In principle, we have already available most of what is required, right? It is just a matter of density, of efficiency, of number of nodes if you wish. When a person understands/remembers something, we all know that it requires a huge amount of actions at a microscopic level about which we don't have a too good understanding. We also know that computers can do virtually everything, but that we need to explain them each single step of the process. You want a computer to distinguish between two pictures? It takes a program of X size. Do you want it to distinguish between two more abstract ideas as defined by a big number of pictures? It takes a program of X^Y size. All this seems quite evident and clear, so why the next logical step seems so difficult to understand? Why expecting a magic setup allowing to easily and immediately come up with a way to restrict that geometrical increase of complexity? On the other hand, if you keep adding layers over and over, step by step until reaching the point where you have converted all our knowledge to a machine-understandable format, it seems pretty clear that we would certainly get a machine able to understand everything as well as the most intelligent person.
In summary, it isn't a matter of how to find the magical way allowing us to avoid the tremendous complexity associated with reliably emulating the human brain, even just a few of its functions. It is a matter of accepting the only thing that you can do to ever be in that scenario. There is no other alternative. Adding simple layers one over the other is all what we know. Perhaps even the human brain works in that way, but much more efficiently. But that issue doesn't even really matter because only know how to do that anyway.
Custom Solvers 2.0 = Alvaro Carballo Garcia = varocarbas.
No syntactical machine can suddenly 'inheret' semantic meaning. All code is syntax save where given meaning by pre-existing semantic entities (like the mind of Man).
AI cannot ever escape its syntactical trap. Science is syntax. Maths is syntax. MEANING (signal vs noise) comes from pre-existing semantic entities (ie., LIFE). What scares betas (but is understood by all scientific/math alphas) is that life could never arise naturally in a CLOCKWORK universe. So life involves another - unknown- syntactical factor.
All science, by definition, must be explained by maths. All maths is limited by the discoveries of Godel and Turing. All maths must be capable of running on a 'Turing' machine- a state machine that allows for ZERO randomness. No mind can exist on a Turing machine.
Alphas (with an understanding of maths) now the absolute implications of the work of Godel and Turing. Half-witted betas who think science "cool" believe in nonsense like 'magic maths' from 'magic' quantum effects. There is no such thing as 'magic' maths- Godel proved this beyond all possible doubt. Quantum effects, being science, are therefore described perfectly by maths- which in its turn musty be able to run on a turing machine- which allows for no 'uncertainty'.
Mind, soul, whatever is an expression of syntax- inherent meaning- and is a concept seperate from what we currently call 'science'. Mind, soul, whatever is an AXIOM- like the specific rules that define the operation of our universe. You cannot, by definition, ask from where axioms derive. They just are. A pure clockwork universe contains NO life. Our universe is where the clockwork is 'contaminated' by something that causes life. Ancient Man called this 'god'- a childish nonsense derived from the Human psychology that is father worship.
AI is the pointless beta rubbish that seeks to deny the principle that life cannot be caused by clockwork. So AI is not real in the sense betas understand it. Alphas know AI is but a BUZZWORD gioven to a particular area of computer programming- that in reality runs to the same principles of all computer programming.
If you believe AI to be 'real', you are no different from those demonic japanese scientists who called their living Human victims 'logs' during WW2 and dissected them while they where still alive and conscious. THIS is why the lie of AI is disseminated- because it is part of the demonisation of the 'other' that allows betas to willingly prepare for war on other Humans. And the ability to get masses of dumb-dumbs (like those of you that believe AI is real) to justify horrors like we see monsters carrying out in Yemen and Gaza is what the game is really about on Slashdot.
Really, AI systems are remarkably stupid. A simple example: tell Google Assistant, or Alexa, NOT, under any circumstances, to give you the weather forecast. They both give you the weather forecast. Their understanding is so incredibly limited that it makes me wonder how much progress has there been, in this respect, within the last half century? What is regrettable (and this article is a breath of fresh air) is that too many in the AI community seem to have forgotten the lessons of history, and are repeating the same mistakes that ended up in the AI Winter. Probably just the first one, for we are likely to be entering another one soon.
Seriously, this can be done. When you create learning algorithms that compete with each other for accuracy and relevance, you kill off those that underperform. Boom, you've added mortality to the learning cycle.
It's called Genetic or Evolutionary Algorithms.
https://towardsdatascience.com/introduction-to-optimization-with-genetic-algorithm-2f5001d9964b
encode a sense, a feeling, a concept? Understanding? Perception? Consciousness?
;)
Seems to me everything today in what is called AI/Machine Learning is little more than (to simplify) a huge case statement/if-elseif/search engine feeding back possible answers. Where the answers themselves must be evaluated, getting back results that in turn need to be evaluated.
Until you are able to encode conceptualization, feeling, understanding and sense of in relation to hard and soft data you may very well just end up in a never ending loop or just a completely wrong answer/response.
How does one create/encode Consciousness?
Just my 2 cents
As the saying goes, AI, like Fusion, has been 10 years away for 30 years now.. and that saying was from 30 years ago.
ML/DL/etc got a lot of people really hopefully since they were SO much easier and you could throw hardware at them, plus they produced great marketing and search results,.. but in many ways we are pretty much where we were in the 70s or 80s in terms of actual AI development when it comes to actual intellegence.
Humans barely get meaning, one shouldn't expect machines to any time soon.
Well, yeah. Throwing stuff at the wall and hoping something words, which is pretty much the core of deep learning/machine learning/etc, is going to have limitations. The main reason the technologies have gotten so popular is that hardware has gotten so much more powerful and thus you can just keep throwing hardware at problems and getting better results out of it without actually developing any understanding of what is happening. These techniques are great for producing answers that don't actually matter, but will always be limited by the lack of needing to actually understand the model.
Modern AI isn't that much different than the AI I learned in school 25 years ago. There are two things that enable AI to be much more useful now, and often seem more powerful than it is:
1) Processing power
2) Dataset size
Both of those are multiple orders of magnitude greater today than 25 years ago, and that is what enables the kind of "flashy" AI that people get to interact with directly. Things like Siri, and photo albums on our phone that can automatically tag images with search terms (like "car", "tree", etc) as well as figure out that the same person is in multiple photos.
Siri, for example, works not because the software is more intelligent and can universally understand English in a speaker-independent manner (like how the human brain would work), but because Apple has the processing power and storage capability to process a voice against dozens (if not hundreds) of speech models in parallel, and then pick the best match. A person with a southern drawl can be understood by Siri because their speech was also matched against a model that was generated from people that speak with a southern drawl. This reminds me that in the 80s I bought a speech recognition IC from Radio Shack for under $10 ( http://21stdigitalhome.blogspo... ). It could only understand around 10 words or so, and had lots of false detections, but it did work. What is the primary difference between it and Siri? Processing power, and the size of the datasets.
So it is has primarily been the physical advancements in computing (processor speed, memory size, storage) that have enabled AI as we know it, and not advancements in the theoretical. To give credit where it is due, certainly some advancements in theory and AI have been made, but not the kinds of breakthroughs that would allow AI to function reasonably on 25 year old hardware, for example.
Better known as 318230.
We think we have understanding. But what does that really mean? That we can recall the linkages clearly to other concepts in our short term memory?
I'm not sure that we can argue that something like Deepmind lacks understanding.
What's missing isn't classification or understanding. What's missing is ability to generalize and learn from way fewer samples.
It makes no sense to say that A.I. fails to work if it arrives at a conclusion that some humans disagree with.
Sure it does - if that conclusion makes it fail at some job. Say you present some 'simple' task, like manipulating objects into some slots they fit in, using some control buttons & a camera feed as input. Some noise in the camera feed: humans aren't bothered much, AI gets confused.
"Fails to do the job at hand" (or succeeds at that job) sounds like a useful metric to me. That's essentially how newly hired personnel is evaluated throughout industry, right? If it looks silly but it works, then who cares - it works. If it's super-advanced but doesn't, then who cares - it's useless. Depending on the job, you may invest some more time to get things to work. Or just give it a few tries & -if unsuccessful- move on to the next applicant. Anything beside that is of academic interest only.
...the long-running ethics violation of the tech/software industry continues. see: http://3seas.org/EAD-RFI-respo...
It should be obvious and in time it will be and then what will be thought of the tech/software industry?
If only we had a system that was designed from the ground up to provide some common sense for AI.
When Fascism comes to America, it will call itself Anti-Fascism, and tell you to give up your guns.
"Recognize speech"
and
"Wreck a nice beach"
can trip up text to speech engines.
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.
AI researchers finally admit that human intelligence cannot be duplicated by machines?
Currently there's a fundamental assumption that awareness of self (and thus intelligence) is the result of the right mix of brain chemistry and electrical impulses, and therefore a silicon-based machine can be just as good as a carbon-based meat machine. But what if this assumption is.... wrong?
Now don't start ranting at me about the nonexistence of Jeebus and Yaweh, yeah I get it, you hate them. But many (probably majority) of you are liberals, and liberals love yoga and buddhism. Every liberal I've met say they admire them. Some of them even practice yoga and buddhist meditation. But you guys are doing just the shallowest, superficial things in yoga and buddhism while missing the most basic and fundamental concept -- you are an immortal spirit that's only inhabiting a physical body. Past lives. Reincarnation. Endless cycle of birth and death. The whole point of buddhism, yoga, hinduism, mysticism, metaphysics, *everything*..... is to break out of this cycle by achieving spiritual enlightenment.
It's like watching a primitive tribesman opening a car door and putting on the seat belt and then getting out and closing the door over and over again, and thinking that's the whole point of a car.
read the defintion before you answer:
https://www.google.com/search?...
gain or acquire knowledge of or skill in (something) by study, experience, or being taught.
"they'd started learning French"
A system that 'only' categorizes , sorts, and manipulates data , does not actually relate to it as representational of the real world, in other words it
still has no 'knowledge' of the objects. They don't ACTUALLY learn they are trained. They no more learn any topic the a parrot learns to talk.
Not to say they aren't extremely useful tools, but to be useful you must keep in mind what a tool is and isn't.
âoeTolerance applies only to persons, but never to truth. Intolerance applies only to truth, but never to persons.
Eh, frankly, for search results, I miss the days when I could get exactly what I searched for. I used to get relevant results on the first page for most of my searches. Sure, they missed a lot of stuff when my searches were too specific, but search mostly worked well. These days, most searches work ok, but there are too many that just come up with a riculous mountain of unrelated garbage that doesnâ(TM)t even have the search terms in it.
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.
Seems we finally have real world verification for Searle's Chinese Room situation. Thank you researchers for finally proving a conjecture from thirty years ago that you continually and blindly ignored. Some of you even argued against it. And now look at the egg on your face.
Ha!
Of course we need better pattern recognition, but I don't think we get there through larger nets and better training. I suspect we've reached a level of training and individual net size that is already adequate. It is actually very surprising that the individual nets we've created can compete as well as they can with humans because they are doing it without the feedback of thousands of other nets that the human brain has.
What is most needed is not better trained specialized nets. We need many nets trained in different things and connected in ways that allow them to correct each other either directly or indirectly through arbitration nets. The feedback from other nets with different specializations would have corrected the interpretation in the examples given because the interpretation didn't make sense.
I've long felt that what we are missing is the OS. We've figured out how to create the subroutines, but little is in place to put the many different subroutines into parallel contact with each other in a useful way.
My gut says this problem is not a big one. As we learn to connect large numbers of nets together in a manner that allows the whole community of nets to settle on a truth and describe it and act on it without any one central thing holding all of the knowledge of the truth, what we think of as understanding will emerge.
The reason current AI technologies will always have a problem mimicking real intelligence is because they thought our brain works like a computer. You have those people conducting countless of researches and projects that attract huge amount of budgets. They keep insisting that our brain works similar to a computer so they could continue receiving more funding, and investors who lack the required knowledge just keep buying into those ideas. Those people would go around and discredit any other counter-hypothesis but the reality is our brain is NOT a computer. On the surface some of its high level functions may appear similar to how a computer operates, e.g store and retrieve memory, information processing etc. but it has been said that most of our behaviors are really driven by our subconsciousness. So far we still don't have access to this layer in our brain and have no idea how it works. How could you simulate an object when you didn't even have an accurate representation of it?
I believe the solution would be to adopt a unambiguous constructed language, such as Lojban, as the primary auxiliary language of the world.
After all, we are limited by the language in which we think and most natural languages are riddled with ambiguities...
It's a cookbook. And since this whole article is merely about extraction of semantic meaning in ambiguous cases, then I will assert that the phrase "As someone who has worked in A.I. for decades" is literally a statement about the matrix and their occupancy within in.
And please could you take a step back because you are pixelating in my vision and I don't like the reminder that you are not real
Some drink at the fountain of knowledge. Others just gargle.
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.
The problem with any machine vision recognition system is that there is really no way of looking at a still image (or even a still scene with depth) and /knowing/ which Objects are discrete and independent
This is not a problem. If you have enough data, the machine will find the patterns.
You are solving the wrong problem with this data abundance. The problem is knowing if the machine learned the right pattern. The surprise is that in many cases it's actually harder to know if the right pattern has been learned than to actually learn the right pattern. Mind bending but there's some quantitative proofs about that called "the no-free-lunch theorem for generalization". The machine will always find a pattern and with enough data it's use of the pattern will defeat your ability to crossvalidate if it learned the right pattern.
One day you will see the queen of hearts and enter a suggestible hypnotic state, put on an elephant costume, and walk across a road and be run down by the Uber that had a pathological classification caused by seeing elephants in the middle of the road.
Some drink at the fountain of knowledge. Others just gargle.
Speech wreck ignition never maid cents two me.
Isn't it absurd to add 'Unartificial' to real intelligence systems?
How do they work? In real life, in all species, there is an element of inherited knowledge. In humans, this is minimal and we must learn from experience and from our mentors. Generally speaking we learn, as all animals and mammals, by experimenting. What doesn't kill us makes us smarter.
We, all of us from microbes to humans, learn by exploring our world without prejudice, in hopes of finding something beneficial to our survival and welfare. When computers can do this they will not be AI, they will be intelligent.
There is one other requirement that biological entities have: reproduction. Each generation of intelligent computers should pave the way to a future generation of more intelligent computers. We should expect them to contribute to the design of the next generation. This is where the singularity leaves us all behind. In the blink of an eye, technology will pass our so-called intelligence and leave us in the dust. It's main advantage is that it will not be encumbered with emotion and loyalty to a nazi in the White House.
...omphaloskepsis often...
In 1972 Hubert L. Dreyfus wrote:
"AI workers, however, want their machines to interact with people in present real-life situations in which objects have special local significance. But computers are not involved in a situation. Every bit of data always has the same value. True, computers are not what Kant would call "transcendentally stupid"; they can apply a rule to a specific case if the specific case is already unambiguously described in terms of general features mentioned in the rule. They can thus simulate one kind of theoretical understanding. But machines lack practical intelligence. They are "existentially" stupid in that they cannot cope with specific situations. Thus they cannot accept ambiguity and the breaking of rules until the rules for dealing with the deviations have been so completely specified that the ambiguity has disappeared. To overcome this disability, AI workers would have to develop an a-temporal, nonlocal, theory of ongoing, situated, human activity."
I keep hearing about machine learning and deep neural nets, but no one talks about the Theory of Practical Activity that would allow AI to be... intelligent. And without that, there is no AI that is going to take my job at the bank, or drive me to the airport, or help me solve a crossword puzzle.
People make the same mistakes. Language is complicated, evolving, and misused constantly. If you told me to type out that sentence, I might assume the guy had a bear for a head also.
I'd like to see a concise definition of what it means for a machine to "understand" something. It's easy to give examples of machines not "understanding" something, but if a machine suddenly dealt with all those examples correctly, could we then say that it "understands" those situations? Or would we find more examples that it gets wrong and say it still doesn't understand?
People are not perfect at interpreting images, either; it's fairly easy to construct an image that a person gets wrong, for instance using forced perspective to make a toy car look like a real one. Do we not "understand"?
Slashdot - News for Nerds, Stuff that Matters, in ISO-8859-1 Has just realised that beta makes this signature redundant
" For example, the uber car that thought the woman in the road was blowing trash."
Good post, but don't spread bullshit about that Uber car.
The car identified the woman as a pedestrian, but was unable to apply the brakes because Uber disabled that part of the software.
I have absolutely no issue with that. But it is not the same thing. It is problem-driven. Much if the AI hype is fantasy driven about the new slaves we are all going to get, or alternatively, the overloads that will kill us. And that is nonsense.
Most ACs are not even worth the keystrokes to insult them. Be generically insulted by this and ignored otherwise.
Indeed. Impressive quantitative advancements, absolutely nothing on the qualitative side. That does extent the applications dramatically, bit these things still have zero (general) intelligence and zero insight.
Most ACs are not even worth the keystrokes to insult them. Be generically insulted by this and ignored otherwise.
The two transcription/translation examples are easy to fix even in dumb AI and totally irrelevant to the question of understanding meaning. A translation engine can easily be taught or learn by training that bareheaded is a word, that "bareheaded" is more frequent than "bear-headed". Also beware: It is not completely impossible that the man could have the head of a bear, and this could be the most likely interpretation if the previous sentence was "I saw two men, one had the head of a lion and the other had the head of a bear", something which a speech-to-text engine could easily recognize without knowing fuck all about meaning. Note that the question of semantics and the question of context can be solved separately; you can assess context in a completely meaning-agnostic engine. Same for the collocation between pig-cochon and pen-porcherie > pen-stylo, something that a translation engine can learn just from looking at training examples.
So AlphaZero becomes exceptionally good at chess with only 4000 TPU-hours of computation, self-playing something on the order of 40 million chess games.
Now just imagine the Mother of AlphaZero where you train an expert system to train up 10 million different AlphaZero-class AIs, so that it can devise the optimal network for any future AI task on pure gut instinct.
A mere 4.5 million TPU years later, and now AI is really cooking with gas.
The AI of the AI remains a little bit out of reach at current computational cost.
When you add "while I was at the farm", google translates correctly to french.
EN: While I was at the farm, I put the pig in the pen.
FR: Pendant que j'étais à la ferme, j'ai mis le cochon dans un enclos.
I actually think it's logical for the AI to only translate like this with the extra context. I mean, are you sure nobody puts pigs in writing tools?
I don't think the mechanism of "meaning" is anything more than a network of associations. The more associations something triggers, the more meaningful it becomes to me. Particularly if the network leads to strong basic emotions, which by themselves are more or less hardwired by evolution.
I don't see why a machine couldn't do all this. As others have pointed out, we train our machines wrong by focusing on quantity rather than quality of data. Also, we don't seem to guide the machine by stating what's essential about the data (as in the example of Russian tank images having grainier quality).
Escher was the first MC and Giger invented the HR department.
It's an old dream, that every person would live a life of leisure with an army of servants.
Technology has significantly reduced the amount of labor we have to do, especially labor at home. This preference makes sense in a way, we don't earn money when we wash our own cloths. But a washing machine manufacturer can to a profit, so they build the machines. We buy the machines because our time can be spent on other things, hopefully more leisure, in practice we end up with households where both husband and wife work outside of the home. The amount of time we save from technology doesn't translate perfectly into an equivalent amount of leisure time.
I think the self-driving car thing is a done deal. It may not be A.I. in the strictest sense. But it does replace a human driver for that narrowly defined task of driving a car. We'll probably accept an incomplete and slightly dangerous implementation rather than wait for a perfectly executed self driving car. Arguments against scenarios where self-driving won't work are already being ignored. We'd need multi-car Uber pile-ups every day for the next few years for the industry to put the brakes on this.
“Common sense is not so common.” — Voltaire
For example, the uber car that thought the woman in the road was blowing trash.
That is not correct. To quote the NTSB investigation:
. As the vehicle and pedestrian paths converged, the self-driving system software classified the pedestrian as an unknown object, as a vehicle, and then as a bicycle with varying expectations of future travel path. At 1.3 seconds before impact, the self-driving system determined that emergency braking was needed to mitigate a collision. According to Uber emergency braking maneuvers are not enabled while the vehicle is under computer control to reduce the potential for erratic vehicle behavior.
The self driving system identified the woman as a bicycle 25 meters away. It identified the woman as a "Vehicle" even further prior. But the system just wasn't sure where the object was headed exactly. The error wasn't in identification the error was that the vehicle should have begun to slow down in case the cyclist was high out of their mind or suicidal to be able to stop if they walked into traffic.
This is a really hard problem for humans too. When driving around the city I try to judge not just where they are headed but also their degree of sobriety. If they are 'obviously' a crazy\high person or inattentive I begin slowing and assume they will just wander across 5 lanes of high speed traffic with no regard for their own safety. If they look sober and attentive then I assume they are going to frogger their way across when it's safe. If they are a bike courier I try to just leave it up to them assuming they'll weave in and out and as long as I travel in a straight line at a constant speed they'll figure it out on their own.
I agree to that, also to the self-driving cars. They are certainly not AGI (Artificial General Intelligence) and anything else is really just dumb automation. But no matter, dumb automation without insights or understanding seems to be perfectly capable of driving a car in all regular situations. What we are actually doing is not building intelligent machines, but finding out that a lot of tasks humans are (so far) needed for do not actually require intelligence. I am perfectly fine with that as well.
I predict that as soon as self-driving cars become generally available, insurance premiums will do the rest very fast, because on average these will cause far less and far less costly accidents. This will probably go even faster in Europe that the US, because here "unlimited" is pretty standard for car insurance and (I think, I have never owned a car) 2M or so is the minimum. This will probably also revolutionize car-sharing, as you can simply order a car to some time and place. Looking forward to that, because I am not a good driver and sometimes a car is useful even in a large city.
Most ACs are not even worth the keystrokes to insult them. Be generically insulted by this and ignored otherwise.
but finding out that a lot of tasks humans are (so far) needed for do not actually require intelligence.
My hope is people don't figure this out before I retire. A lot of what goes on in software engineering is repetitive, formulaic, and un-creative.
I predict that as soon as self-driving cars become generally available, insurance premiums will do the rest very fast, because on average these will cause far less and far less costly accidents.
There is a curious dance going on right now between car manufacturers, self-driving car systems developers, and insurance companies on how to lobby the government for where to assign responsibility for self-driving car accidents. Individuals don't have any lobbyist so they'll probably get the short end of the stick on this one.
Looking forward to that, because I am not a good driver and sometimes a car is useful even in a large city.
You're a rare breed, most people insist they are above average drivers. (heh)
“Common sense is not so common.” — Voltaire