Religious people have a wonderful ability to interpret their stated religion according to their own values. For example, The majority of the citizens of France in 2013 were Catholic; that's the year France legally recognized same sex marriage. Opinion polls in India put the issue at a three way tie, yes/no/don't know.
In regular co-evolution you'd generally have a cost/fitness function to judge fitness by. Evolutionary algorithms are essentially optimizers that do well with non-convex problems.
GANs were invented to answer the question "what do you do when you don't know what the cost function is?" in the particular circumstance when you want to build a generative model. You want the output of your generative model to be as much as possible like a true example, but it's difficult to properly measure similarity. There are lots of similarity metrics, but they all tend to concentrate on particular features, and don't work terribly well for generation.
So in a GAN you let a second neural network *learn* what the cost function is, based on the simple objective: decide which of these examples are real and which were generated. You train both models together with the idea that the generative model has to figure out how to create examples that are enough like the real ones that the critic can't find a way to tell them apart.
As you might imagine, the whole thing depends on the capacity of the critic, and GANs have been quite difficult to train because the generator is very good at exploiting limitations in the the critic.
IIRC, the Lion Air flight did put the flaps back down at one point, and everything settled down, until they put them back up again. The Ethiopian plane was apparently too fast, although presumably there was a window of opportunity earlier.
Yes. Which is apparently in Boeing's software fix. The solution is not three sensors (of the same type) as was suggested by the OP. Two AOA sensors are fine.
See what I did there? Same thing you did: assert an opinion with nothing to back it up. See the post you replied to for said backup. If you'd like to actually reply to that, instead of ranting, happy to discuss.
Two sensors is fine, because it isn't really a safety critical system. All engine-under-wing airliners have the same instability problem in a stall, and the MAX isn't that much worse. If the system turned off on failure, it would be fine.
Whoever approved using only one sensor was an idiot.
Hey, you know those things that let you carbonate your own water? You know how soda water has that distinct taste? The taste of carbonic acid? You should drop a pH test strip in some of that stuff.
That means Russia and Canada will lose their ability to grow wheat where they currently do.
Highly doubtful.... the prediction was that winds from the interior of the US would dominate causing less rainfall and longer and more severe droughts.
Basically, yes. Many computer vision systems don't have the background knowledge that we take for granted. There are some interesting papers showing that a lot of the apparent mismatch between human performance and some of these systems can be attributed to background concepts. In one, they designed a game that humans could easily master with no instructions, but their AI system had great difficultly with. It had elements like a pink princess in a pointy hat at the top, and the hero (a blue icon) had to jump up and rescue her. As they removed these stereotypical clues, the human performance degraded, eventually to the level of the AI system. What you describe is along the same lines.
As for #1, there are lots of videos of trucks hitting overpasses on the Internet. In a few cases people have set up permanent webcams because it happens so often.
This stuff is all very new. People know about adversarial attacks, but they don't necessarily train against them. The Tesla autopilot in particular is very specifically a driver aid, it's not supposed to be foolproof, and isn't subject to any of the rigorous engineering requirements that a non-supervised system would be.
Training against adversarial attacks doesn't mean training against a particular set of stickers or whatever. It involves hardening the whole system to attacks, like this one, that involve subtle, almost undetectable, alterations to the input.
I'm not sure how a news report about temperatures between -20 and -27 is some kind of proof against my statement "it's been years since we had a good run of -50C."
It has been a while since I took formal logic though. Care to explain?
I think humans give themselves way too much credit for their "thinking" ability. The research suggests what we do is nothing like the logical reasoning most of us assume. Mysterious processes tell us the answer and then, if pressed, we justify it to ourselves.
I agree, and I think people who use stories like this as "oh look, proof self-driving cars will never work" are wrong. But adversarial examples are an issue that should be solved. I don't think it's a terribly difficult solution though. One of the great things about adversarial examples is that you don't even need more training data to get started, just the output of your own adversarial generator.
The difference being a human that sees lane markers leading into active oncoming traffic will decide there are shenigans and not follow.
I've seen lots of drivers do exactly this. That was in Montreal though, so it may be significantly less common elsewhere.
It points to a big gap in machine learning strategies in general: Training generally happens focused on positive correlations and not a lot of injection of maliciously designed data.
There's usually lots of negative data in training sets, but you're right, not so much malicious. But the real human-type armour against malicious data is noise and experience. The way these adversarial attacks work is to compute the steepest path gradient away from the right answer and design a pattern that exploits it. By the time a human is allowed to drive, it's brain has sifted through petabytes of noisy imaging data, so most of those quick, easy adversarial paths have been closed through random chance. Some are reinforced though, and we call those optical illusions.
Hm... 3d scan of the relevant anatomies, and matching based on topological compatibility. It's even more fun for the woman, so attracts the right demographic.
By "weight" I think most people, including, I bet, the two statistical anomalies, actually mean proportions of a woman's figure.
People (not just guys) like to assess themselves and others with something akin to body mass index, which depends on height and weight. What they *mean* by that clumsy metric is waist to hip ratio. Waist to hip ratio also works better than BMI for predicting health. Which is exactly why it correlates with our subjective judgement of attractiveness.
Slashdot Mirror
Religious people have a wonderful ability to interpret their stated religion according to their own values. For example, The majority of the citizens of France in 2013 were Catholic; that's the year France legally recognized same sex marriage. Opinion polls in India put the issue at a three way tie, yes/no/don't know.
In regular co-evolution you'd generally have a cost/fitness function to judge fitness by. Evolutionary algorithms are essentially optimizers that do well with non-convex problems.
GANs were invented to answer the question "what do you do when you don't know what the cost function is?" in the particular circumstance when you want to build a generative model. You want the output of your generative model to be as much as possible like a true example, but it's difficult to properly measure similarity. There are lots of similarity metrics, but they all tend to concentrate on particular features, and don't work terribly well for generation.
So in a GAN you let a second neural network *learn* what the cost function is, based on the simple objective: decide which of these examples are real and which were generated. You train both models together with the idea that the generative model has to figure out how to create examples that are enough like the real ones that the critic can't find a way to tell them apart.
As you might imagine, the whole thing depends on the capacity of the critic, and GANs have been quite difficult to train because the generator is very good at exploiting limitations in the the critic.
IIRC, the Lion Air flight did put the flaps back down at one point, and everything settled down, until they put them back up again. The Ethiopian plane was apparently too fast, although presumably there was a window of opportunity earlier.
Yes. Which is apparently in Boeing's software fix. The solution is not three sensors (of the same type) as was suggested by the OP. Two AOA sensors are fine.
Soooo... you don't use an IDE?
I'm not making fun of you. I code in a minimal text editor. Code completion drives me up the wall. But almost nobody else does that.
Two sensors is fine.
See what I did there? Same thing you did: assert an opinion with nothing to back it up. See the post you replied to for said backup. If you'd like to actually reply to that, instead of ranting, happy to discuss.
Two sensors is fine, because it isn't really a safety critical system. All engine-under-wing airliners have the same instability problem in a stall, and the MAX isn't that much worse. If the system turned off on failure, it would be fine.
Whoever approved using only one sensor was an idiot.
Both flights seem to have run into trouble when they put their flaps up, engaging the normal (non-takeoff) flight regime.
Hey, you know those things that let you carbonate your own water? You know how soda water has that distinct taste? The taste of carbonic acid? You should drop a pH test strip in some of that stuff.
Hm....
Basically, yes. Many computer vision systems don't have the background knowledge that we take for granted. There are some interesting papers showing that a lot of the apparent mismatch between human performance and some of these systems can be attributed to background concepts. In one, they designed a game that humans could easily master with no instructions, but their AI system had great difficultly with. It had elements like a pink princess in a pointy hat at the top, and the hero (a blue icon) had to jump up and rescue her. As they removed these stereotypical clues, the human performance degraded, eventually to the level of the AI system. What you describe is along the same lines.
As for #1, there are lots of videos of trucks hitting overpasses on the Internet. In a few cases people have set up permanent webcams because it happens so often.
You're kind of a disagreeable human, hey? Or maybe a chatbot that's been listening to Reddit for too long?
This stuff is all very new. People know about adversarial attacks, but they don't necessarily train against them. The Tesla autopilot in particular is very specifically a driver aid, it's not supposed to be foolproof, and isn't subject to any of the rigorous engineering requirements that a non-supervised system would be.
Training against adversarial attacks doesn't mean training against a particular set of stickers or whatever. It involves hardening the whole system to attacks, like this one, that involve subtle, almost undetectable, alterations to the input.
I didn't say I was from Edmonton, nor a major (or any other kind of) city.
You know what they say about assumptions.
I'm not sure how a news report about temperatures between -20 and -27 is some kind of proof against my statement "it's been years since we had a good run of -50C."
It has been a while since I took formal logic though. Care to explain?
I think humans give themselves way too much credit for their "thinking" ability. The research suggests what we do is nothing like the logical reasoning most of us assume. Mysterious processes tell us the answer and then, if pressed, we justify it to ourselves.
I agree, and I think people who use stories like this as "oh look, proof self-driving cars will never work" are wrong. But adversarial examples are an issue that should be solved. I don't think it's a terribly difficult solution though. One of the great things about adversarial examples is that you don't even need more training data to get started, just the output of your own adversarial generator.
I've seen lots of drivers do exactly this. That was in Montreal though, so it may be significantly less common elsewhere.
There's usually lots of negative data in training sets, but you're right, not so much malicious. But the real human-type armour against malicious data is noise and experience. The way these adversarial attacks work is to compute the steepest path gradient away from the right answer and design a pattern that exploits it. By the time a human is allowed to drive, it's brain has sifted through petabytes of noisy imaging data, so most of those quick, easy adversarial paths have been closed through random chance. Some are reinforced though, and we call those optical illusions.
It makes them a lot milder though. It's been years since we had a good run of -50C.
The greatest warming is in the middle of nowhere. I'm from the northern Canadian prairies. There isn't any asphalt.
It's neither. They engineered yeast to produce it.
There are restaurants that make burgers the same way they make things like tartare. You can eat those rare.
You should never eat random ground meat you get at the supermarket anything other than cooked through.
Hm... 3d scan of the relevant anatomies, and matching based on topological compatibility. It's even more fun for the woman, so attracts the right demographic.
It's true. A LOT of women put a height requirement in their dating profile. I doubt most guys would dare put a weight requirement in.
On the other hand, I've been told by a lot of female friends that a surprising number of guys have strong height preferences for women too.
By "weight" I think most people, including, I bet, the two statistical anomalies, actually mean proportions of a woman's figure.
People (not just guys) like to assess themselves and others with something akin to body mass index, which depends on height and weight. What they *mean* by that clumsy metric is waist to hip ratio. Waist to hip ratio also works better than BMI for predicting health. Which is exactly why it correlates with our subjective judgement of attractiveness.