In the 70s the idea was to program a bunch of rules into a computer and, if you had the right rules, presto, it would be intelligent. That approach is vulnerable to the "we're not smart enough to program an AI" objection.
The vast majority of modern attempts at AI, including neural networks, learn. They're not programmed. Just like a person. Other approaches use hardware, including natural neurons grown on an artificial substrate, and etched artificial "neurons."
If you want to define "algorithmic approach" to mean any kind of algorithm at all, then your argument is easily countered by observing that the laws of physics are an algorithm and we somehow manage to exist.
No. We build intelligent brains all the time. We're just talking about making a few optimizations, possibly moving to a different process. We (and the universe) build galaxies far less often.
"Even IBM's Watson, supposedly the pinnacle of AI technology, couldn't understand what Alex Trebek was saying. Yup, somebody typed in the question on a keyboard beforehand and fed it to the computer."
Did you see the story today where Microsoft just demonstrated live translation over Skype? Machines can understand Alex Trebek now. Even if he's speaking German.
Because we don't really understand how intelligence works, it's foolish to try to predict when we'll duplicate it. Equally foolish to say that we won't have it in the next 50 years as to say we will have it in the next 5.
Are you serious? YOU are an example of intelligence in this universe. If nothing else, we can either grow neurons on chips (yes, we can do that now) or fiddle with existing human brains (we can also do that). AI is one of the last frontiers where we're trying to duplicate something that we have a common example of in nature.
Who cares about "directed fashion"? There are currently efforts at both making machines smarter and making people smarter. Neither one is particularly "directed." Software AI research is dominated by approaches that set up comparatively simple frameworks, some rules, and rely on self-organization. Hardware research involves growing or making things that are or resemble neurons, letting them hook themselves up, and seeing what happens. Biological approaches tinker with existing functions to see if poking the right thing improves them.
The "programming an AI" approach was (correctly, I believe) seen to be hopeless between the sixties and nineties and has been essentially abandoned.
Suppose I want a measure of your knowledge of world history. If I ask you all questions about American history you may do very well (or very poorly). The same if I ask you all questions about south east Asian history. Only by asking you questions about different time periods and different regions can I get a comprehensive estimate of your average knowledge of world history.
Most measurements of that type are multi-factorial. Scientific literacy is no exception. According to Wikipedia's definition (http://en.wikipedia.org/wiki/Scientific_literacy), it's supposed to depend on a variety of things.
One of the big criticisms of most IQ tests is that they are culturally biased. The questions are such that people from certain cultures have an advantage, regardless of their IQ. My first two example historical knowledge tests were also culturally biased. The way to avoid that is to ask a variety of questions. Your knowledge of SE Asian history may not correlate well with your knowledge of American history. That's good. It seems likely that there is a similar cultural/relgious bias operating here. People from particular religions that object to evolution (e.g. evangelical US sects) may score poorly on this question while doing well elsewhere. Other religions (e.g. catholicism, which officially supports evolution) may do well on this question, but potentially poorly on others that happen to conflict with their beliefs.
What data? There is no gold standard measure of "scientific literacy". The *point* is to make one. Again, you're misrepresenting what was done.
You can argue that a willingness to believe in magic over scientific evidence (what the evolution question is testing) is not relevant to scientific literacy and questions testing that factor should be eliminated from the test, but you can't point to one question, say it doesn't correlate with the others, and discard it on that basis. That's the *opposite* of what you want to do. A well designed testing instrument will have questions that are all at least partially independent in order to properly cover the parameter space. A few redundancies are good as a check, but if many of your questions are redundant you need to eliminate *those* to shorten the test.
If you want a measure of scientific *knowledge* your questions should be along the lines of "are you familiar with the theory of evolution / gravity / fluid mechanics?" or "does the theory of gravity predict that all matter attracts all other matter?". Do you believe in" is explicitly testing whether the person accepts the scientific results over whatever personal beliefs they hold. That is, it's testing whether they accept that science, which is a process not a body of knowledge, is the best available means for generating knowledge. I said scientific literacy is a stupid term, but according to wikipedia (http://en.wikipedia.org/wiki/Scientific_literacy) it is not a measure just of scientific knowledge. It's supposed to include several things, such as an ability to "evaluate the quality of scientific information ", "express positions that are scientifically and technologically informed", and "describe, explain, and predict natural phenomena", all of which are compromised if your willingness to accept scientific results is subject to them not conflicting with your personal beliefs.
That is, according to the definition of scientific literacy, questions that test a person's willingness to disbelieve scientific results based on unsupported personal or cultural beliefs are *required* for a good test of scientific literacy. If they don't correlate with other questions on the test, so much the better - they're capturing something that the rest of your test is not.
No, you've misrepresented the data. Right in the summary:
"People who score at or near the top on the remaining portions of the test aren't any more likely to get this item "correct" than those who do poorly on the remaining portions."
"What the NSF's evolution item does measure, researchers have concluded, is test takers' cultural identities, and in particular the centrality of religion in their lives."
They're trying to measure "scientific literacy" (which is a stupid term). The answers to the evolution question don't correlate with the answers to the other questions because it's measuring something different. They've concluded it's measuring people's inclination to believe in religion, presumably over science. That would seem to be an important factor in scientific literacy, so the evolution question is actually capturing something that is missed by the other questions.
The point of a multi-question test is that the questions should measure different things - they won't all necessarily correlate well with each other. Measuring someone's inclination to believe religion over science would seem to be a valuable part of assessing their scientific literacy.
That's pretty much the way deep networks work. You show them a bunch of things and they generate internal representations of them. Cats and dogs are spontaneously divided into different representations, which can then be translated into class labels.
From the description in the article, it sounds like what's happening here is that the researchers have purposely designed test cases that confuse the network. It's less likely to happen with common things in people because we have much more thorough training - before you can reliably recognize cats you've seen a LOT of them, from all angles. Also, your eyes are constantly shifting their viewpoint a bit so that they get jittered images. They might be able to reduce this problem a lot by training the network on variations of the same images randomly shifted by a pixel or two.
It's also not particularly hard to come up with images that confuse the human brain.
The evil "left as an exercise for the reader" part of textbooks where the author shows you a bunch of examples then gives you a problem that's related to those, but just enough different in some small way that it's fiendishly difficult. Or, more generally, the trick question.
That's not really true. ANNs, particularly deep ones, share a lot of features with specific parts of the brain, especially primary sensory processing areas. It's quite reasonable to ask whether emergent properties of deep AANs are also found in the analogous systems in the brain (and vice versa). Some, such as visual processing done by simple, then complex cells, are already known. An ANN isn't a simulation of any part of the brain, but it's an analogous system that does share some properties and might well share others.
Your knowledge is out of date. Support vector machines can replace shallow neural networks. The deep ones have serious, mathematically proven, advantages over shallow AANs and SVMs.
If you were taking a machine learning class a year ago that said nobody is using AANs then it was five to ten years out of date. Google has put quite a few resources into them, including buying (er, hiring) one of the pioneers of deep networks.
Reviewers reject radical ideas that are insufficiently supported by evidence. Yes, the more radical the idea, the more evidence it needs to support it.
Contrary to popular belief, "I think ${crazy_idea}" isn't enough to get a paper published. Except in arts journals. And medicine.
A test not producing results that reflect what it purports to be measuring is kinda the definition of inaccurate. Perhaps you're confusing accuracy and precision?
In Canada most places use lined glasses. The bartender is almost never going to actually look at the line, so you always get a bit extra. The point is that you know you're going to get a minimum amount of beer. Crappy bartender who pours a giant amount of head? No problem.
Fuel economy is the same thing. You can't tell if the inflated numbers for one car are more or less inflated than the numbers for another car. Not to mention you can't actually do any planning based on them.
An inaccurate but precise measure is great if it's consistently inaccurate. But if it's consistently inaccurate, why not just measure the inaccuracy and correct all the values?
Because inaccurate measurements are rarely consistently so.
Pinene is an insecticide, and is potentially dangerous. Peppermint additives to things like shampoo have been certified as safe, provided the pinene content is low enough. Some varieties of mint produce enough pinene to not be safe. You don't want to be putting those on your head, or anywhere else.
Chemophobia isn't healthy, particularly when it's combined with naturophillia.
Most animals that we haven't domesticated are very difficult to domesticate. Many don't breed well in captivity, require specialized diets, etc. Even something as simple as zebras having a knack for avoiding lassos meant they were never ridden or used as beasts of burden, except in special, limited, usually ceremonial circumstances.
Slashdot Mirror
In the 70s the idea was to program a bunch of rules into a computer and, if you had the right rules, presto, it would be intelligent. That approach is vulnerable to the "we're not smart enough to program an AI" objection.
The vast majority of modern attempts at AI, including neural networks, learn. They're not programmed. Just like a person. Other approaches use hardware, including natural neurons grown on an artificial substrate, and etched artificial "neurons."
If you want to define "algorithmic approach" to mean any kind of algorithm at all, then your argument is easily countered by observing that the laws of physics are an algorithm and we somehow manage to exist.
That's ridiculous. http://www.cs.columbia.edu/~se...
Processor speeds have, and continue to, increase exponentially.
Most people gave up on algorithmic design for AI in the seventies. It doesn't even factor into current approaches.
No. We build intelligent brains all the time. We're just talking about making a few optimizations, possibly moving to a different process. We (and the universe) build galaxies far less often.
"Even IBM's Watson, supposedly the pinnacle of AI technology, couldn't understand what Alex Trebek was saying. Yup, somebody typed in the question on a keyboard beforehand and fed it to the computer."
Did you see the story today where Microsoft just demonstrated live translation over Skype? Machines can understand Alex Trebek now. Even if he's speaking German.
Because we don't really understand how intelligence works, it's foolish to try to predict when we'll duplicate it. Equally foolish to say that we won't have it in the next 50 years as to say we will have it in the next 5.
Are you serious? YOU are an example of intelligence in this universe. If nothing else, we can either grow neurons on chips (yes, we can do that now) or fiddle with existing human brains (we can also do that). AI is one of the last frontiers where we're trying to duplicate something that we have a common example of in nature.
Why not? Any pair of opposite sex idiots can make an intelligence and teach it to be reasonably functional, and to make more intelligences.
Who cares about "directed fashion"? There are currently efforts at both making machines smarter and making people smarter. Neither one is particularly "directed." Software AI research is dominated by approaches that set up comparatively simple frameworks, some rules, and rely on self-organization. Hardware research involves growing or making things that are or resemble neurons, letting them hook themselves up, and seeing what happens. Biological approaches tinker with existing functions to see if poking the right thing improves them.
The "programming an AI" approach was (correctly, I believe) seen to be hopeless between the sixties and nineties and has been essentially abandoned.
You're setting up a false dichotomy.
Suppose I want a measure of your knowledge of world history. If I ask you all questions about American history you may do very well (or very poorly). The same if I ask you all questions about south east Asian history. Only by asking you questions about different time periods and different regions can I get a comprehensive estimate of your average knowledge of world history.
Most measurements of that type are multi-factorial. Scientific literacy is no exception. According to Wikipedia's definition (http://en.wikipedia.org/wiki/Scientific_literacy), it's supposed to depend on a variety of things.
One of the big criticisms of most IQ tests is that they are culturally biased. The questions are such that people from certain cultures have an advantage, regardless of their IQ. My first two example historical knowledge tests were also culturally biased. The way to avoid that is to ask a variety of questions. Your knowledge of SE Asian history may not correlate well with your knowledge of American history. That's good. It seems likely that there is a similar cultural/relgious bias operating here. People from particular religions that object to evolution (e.g. evangelical US sects) may score poorly on this question while doing well elsewhere. Other religions (e.g. catholicism, which officially supports evolution) may do well on this question, but potentially poorly on others that happen to conflict with their beliefs.
What data? There is no gold standard measure of "scientific literacy". The *point* is to make one. Again, you're misrepresenting what was done.
You can argue that a willingness to believe in magic over scientific evidence (what the evolution question is testing) is not relevant to scientific literacy and questions testing that factor should be eliminated from the test, but you can't point to one question, say it doesn't correlate with the others, and discard it on that basis. That's the *opposite* of what you want to do. A well designed testing instrument will have questions that are all at least partially independent in order to properly cover the parameter space. A few redundancies are good as a check, but if many of your questions are redundant you need to eliminate *those* to shorten the test.
If you want a measure of scientific *knowledge* your questions should be along the lines of "are you familiar with the theory of evolution / gravity / fluid mechanics?" or "does the theory of gravity predict that all matter attracts all other matter?". Do you believe in" is explicitly testing whether the person accepts the scientific results over whatever personal beliefs they hold. That is, it's testing whether they accept that science, which is a process not a body of knowledge, is the best available means for generating knowledge. I said scientific literacy is a stupid term, but according to wikipedia (http://en.wikipedia.org/wiki/Scientific_literacy) it is not a measure just of scientific knowledge. It's supposed to include several things, such as an ability to "evaluate the quality of scientific information ", "express positions that are scientifically and technologically informed", and "describe, explain, and predict natural phenomena", all of which are compromised if your willingness to accept scientific results is subject to them not conflicting with your personal beliefs.
That is, according to the definition of scientific literacy, questions that test a person's willingness to disbelieve scientific results based on unsupported personal or cultural beliefs are *required* for a good test of scientific literacy. If they don't correlate with other questions on the test, so much the better - they're capturing something that the rest of your test is not.
No, you've misrepresented the data. Right in the summary:
"People who score at or near the top on the remaining portions of the test aren't any more likely to get this item "correct" than those who do poorly on the remaining portions."
"What the NSF's evolution item does measure, researchers have concluded, is test takers' cultural identities, and in particular the centrality of religion in their lives."
They're trying to measure "scientific literacy" (which is a stupid term). The answers to the evolution question don't correlate with the answers to the other questions because it's measuring something different. They've concluded it's measuring people's inclination to believe in religion, presumably over science. That would seem to be an important factor in scientific literacy, so the evolution question is actually capturing something that is missed by the other questions.
The point of a multi-question test is that the questions should measure different things - they won't all necessarily correlate well with each other. Measuring someone's inclination to believe religion over science would seem to be a valuable part of assessing their scientific literacy.
That's pretty much the way deep networks work. You show them a bunch of things and they generate internal representations of them. Cats and dogs are spontaneously divided into different representations, which can then be translated into class labels.
From the description in the article, it sounds like what's happening here is that the researchers have purposely designed test cases that confuse the network. It's less likely to happen with common things in people because we have much more thorough training - before you can reliably recognize cats you've seen a LOT of them, from all angles. Also, your eyes are constantly shifting their viewpoint a bit so that they get jittered images. They might be able to reduce this problem a lot by training the network on variations of the same images randomly shifted by a pixel or two.
It's also not particularly hard to come up with images that confuse the human brain.
Yeah right. Brains make mistakes all the time, but natural selection has tuned them to err on the side of paranoia.
The evil "left as an exercise for the reader" part of textbooks where the author shows you a bunch of examples then gives you a problem that's related to those, but just enough different in some small way that it's fiendishly difficult. Or, more generally, the trick question.
That's not really true. ANNs, particularly deep ones, share a lot of features with specific parts of the brain, especially primary sensory processing areas. It's quite reasonable to ask whether emergent properties of deep AANs are also found in the analogous systems in the brain (and vice versa). Some, such as visual processing done by simple, then complex cells, are already known. An ANN isn't a simulation of any part of the brain, but it's an analogous system that does share some properties and might well share others.
Deep networks automatically learn to recognize image elements. That's one of their most interesting features.
Your knowledge is out of date. Support vector machines can replace shallow neural networks. The deep ones have serious, mathematically proven, advantages over shallow AANs and SVMs.
If you were taking a machine learning class a year ago that said nobody is using AANs then it was five to ten years out of date. Google has put quite a few resources into them, including buying (er, hiring) one of the pioneers of deep networks.
Reviewers reject radical ideas that are insufficiently supported by evidence. Yes, the more radical the idea, the more evidence it needs to support it.
Contrary to popular belief, "I think ${crazy_idea}" isn't enough to get a paper published. Except in arts journals. And medicine.
A test not producing results that reflect what it purports to be measuring is kinda the definition of inaccurate. Perhaps you're confusing accuracy and precision?
In Canada most places use lined glasses. The bartender is almost never going to actually look at the line, so you always get a bit extra. The point is that you know you're going to get a minimum amount of beer. Crappy bartender who pours a giant amount of head? No problem.
Fuel economy is the same thing. You can't tell if the inflated numbers for one car are more or less inflated than the numbers for another car. Not to mention you can't actually do any planning based on them.
An inaccurate but precise measure is great if it's consistently inaccurate. But if it's consistently inaccurate, why not just measure the inaccuracy and correct all the values?
Because inaccurate measurements are rarely consistently so.
Ignorance encourages mass generalization.
Enjoy your pesticide laden shampoo, and your fear of OMG toxic!
Pinene is an insecticide, and is potentially dangerous. Peppermint additives to things like shampoo have been certified as safe, provided the pinene content is low enough. Some varieties of mint produce enough pinene to not be safe. You don't want to be putting those on your head, or anywhere else.
Chemophobia isn't healthy, particularly when it's combined with naturophillia.
Most animals that we haven't domesticated are very difficult to domesticate. Many don't breed well in captivity, require specialized diets, etc. Even something as simple as zebras having a knack for avoiding lassos meant they were never ridden or used as beasts of burden, except in special, limited, usually ceremonial circumstances.