"Schneider and colleagues' algorithm allows for occasional reverse steps that can unlock better solutions"... "The algorithm uses backward moves often at the start of a packing process but they become less frequent as it closes in on the final solution"
That could describe just about every packing algorithm ever tried, ever. Can't we get a well-written article that actually says why it's different?
These things keep coming up, therefore I must keep repeating myself. Every time any sort of "screening" for terrorists comes up, I point out the problem of overfitting. This comes into play in two ways:
1) It's a simple mathematical fact that if you do not have a suitably large training set (i.e. *actual terrorists*) to study while developing your model, then your model is 100% horse crap, no matter how many anecdotes or pet theories you put behind it.
2) Even if you do have enough *actual terrorists* to study, there is no guarantee that your model actually predicts terrorists when applied to new subjects. It is more likely that your model is latching on to some coincidental pattern in your training data, then actually predicting anything. The problem gets worse as you add more parameters to your model. To reduce overfitting, you need a very large sample, and you need more samples to test your model after the fact. Even then, overfitting can still occur.
Fact: We *do not* have a large sample of actual terrorists that we can freely study for developing a model for predicting terrorist behavior.
Inescapable conclusion: This will result in a *huge number of false positives*.
Additional problem: The *actual terrorists* can learn your (flawed) model by performing their own tests on your system, and by choosing people who routinely get through your system, they can *increase their success rate*. This was proven in a paper a few years ago.
Oh well, here's to tons of wasteful spending that will inconvenience (and sometimes physically harm) honest people and increase the likelihood of successful terrorism!
*Any* business based solely on patents is pure overhead for the rest of us. Overhead of the patent system, I mean. The fact that you can have a growth business of just dealing with patents and nothing else means that the overhead of the patent system itself grows over time. At some point the overhead costs more than the benefit that we get from the system! Don't get me started on insurance companies, lawyers, bankers and brokers either. Each system has players that are pure overhead, and increasingly so.
Finally we will be treated to live-action adaptations of such classics as Buck Rogers, Sherlock Holmes, Wonder Woman, Red Sonja, Barbarella, The Green Hornet and The Flash! Wait, what? Those were made into live-action years ago? Oh bloody hell. Oh well, I for one am looking forward to the sequels to recent films in that list, such as Iron Man and (maybe sequel) Superman. I guess I will treat the rest as not-really-sequels and / or pseudo-remakes. Lots of movies were better the second time around anyway. Most aren't, but here's hoping.
So child aggression before video games existed was caused by... what again? Oh, TV, right? And before that it was caused by... comics, right? And before that there was no child aggression.
Will someone explain to me how a market can be truely free and not devolve into a conglomeration of companies screwing the populous out of their money through treatury and customer lock-in?
It can't. A free market will usually evolve into a cartel. The theory that a free market is good for everybody relies on the assumption that everyone involved has access to enough information to make informed decisions. This does not happen in real life, and in fact the cartels will try to hide the information that you need, unless you have some independent agency operating as you suggested. End of story.
I said it before and I'll say it again:
Any model that is built on 10 or 20 positive examples from a population of 6,000,000,000 is going to suffer from overfitting. Not just a little overfitting... I mean it's going to overfit like a mo-fo. There's just no way, and I mean NO way, to create a statistically significant test based on the data we have on who is and who is not an ACTUAL terrorist.
NOT old news. Google for "2008 siggraph papers". Read the paper. Google for "2004 siggraph papers". Read about the old paper. Note the differences. Tim Rowley posts links to the papers from each year, so his site is recommended. Virtually all of these image-processing-related news items can be read long before they reach slashdot simply by keeping up with the latest papers from siggraph.
In case you're lazy, the old paper is "Non-photorealistic Camera: Depth Edge Detection and Stylized Rendering Using a Multi-Flash Camera". Oddly, it's offline now. But I do have a copy of it on my hard drive. If you're not lazy, I HIGHLY recommend perusing all of the years' papers listed on Tim's site.
No, I would say the best way geeks can save the planet is to engineer a way to clean up the entire environment, as well as engineering a new energy infrastructure that needs less clean-up maintenance. Everything else is secondary to that.
But what do you patent? The thing you built? Someone else could build nearly the same thing without paying any royalties. The issue lies in how you define "nearly the same". Should it be the same on the quantum level? On a macroscopic level? The same components in the same relative configuration? In software, should it be the same machine code? The same source code? A logically equivalent algorithm? A similar user experience? This needs to be defined or you have nothing.
I called this when the PhysX cards first came out. I told my excited coworkers, "these cards are going to be irrelevant pretty soon, because it will all move to the GPU". They looked at me funny.
Hahaha, that's awesome. It's like the psychologists with the chimps choosing different colored M&Ms. Turns out sometimes you need to do math when you do statistics!
Any model that is built on 10 or 20 positive examples from a population of 6,000,000,000 is going to suffer from overfitting. Not just a little overfitting... I mean it's going to overfit like a mo-fo. There's just no way, and I mean NO way, to create a statistically significant test based on the data we have on who is and who is not an ACTUAL terrorist. Books like this are pure speculation.
And which test are you talking about, that gets more and more accurate the more times you apply it to an individual? Can you apply it to the same person 10000 times in one sitting, to get the required 99.9999% accuracy that it would need to be useful? Or do you need to invite them back to take the test on 10000 different occasions? Enlighten us more.
You should read the competition rules. The test set is so enormous that you would need 2^something_huge entries to see the results we've seen based on randomness. I did a back-of-the-envelope calculation at the beginning of the competition to see if a random search would be feasible to win the prize, and it's not. Not in a million years. Literally.
Ray Tracing will still be slow on global illumination anyway. The more reflections you have the longer it takes, so it's not going to look as good too. Look as good as what? The magical non-raytracing global illumination algorithm that you are hiding from the rest of the world? I don't think "global illumination" means what you think it means. Reality check: ray tracing methods are the only way to correctly compute GI today. And by "correctly", I mean "does not fail miserably on the first non-trivial case encountered". Look up Ingo Wald's work, and Eric Veach while you're at it. If you can do it better, then do us a favor and write it up in a SIGGRAPH paper.
If you can't build a large hadron supercollider in the game and get new insights into particle physics, in real time, then it fails the test. This is NOT near future.
Slashdot Mirror
"Schneider and colleagues' algorithm allows for occasional reverse steps that can unlock better solutions" ... "The algorithm uses backward moves often at the start of a packing process but they become less frequent as it closes in on the final solution"
That could describe just about every packing algorithm ever tried, ever. Can't we get a well-written article that actually says why it's different?
I'm sorry, what? Are they complaining that it's hard to make money because there is competition? Hahahahaha. HAHAHAhahahahaha.
These things keep coming up, therefore I must keep repeating myself. Every time any sort of "screening" for terrorists comes up, I point out the problem of overfitting. This comes into play in two ways:
1) It's a simple mathematical fact that if you do not have a suitably large training set (i.e. *actual terrorists*) to study while developing your model, then your model is 100% horse crap, no matter how many anecdotes or pet theories you put behind it.
2) Even if you do have enough *actual terrorists* to study, there is no guarantee that your model actually predicts terrorists when applied to new subjects. It is more likely that your model is latching on to some coincidental pattern in your training data, then actually predicting anything. The problem gets worse as you add more parameters to your model. To reduce overfitting, you need a very large sample, and you need more samples to test your model after the fact. Even then, overfitting can still occur.
Fact: We *do not* have a large sample of actual terrorists that we can freely study for developing a model for predicting terrorist behavior.
Inescapable conclusion: This will result in a *huge number of false positives*.
Additional problem: The *actual terrorists* can learn your (flawed) model by performing their own tests on your system, and by choosing people who routinely get through your system, they can *increase their success rate*. This was proven in a paper a few years ago.
Oh well, here's to tons of wasteful spending that will inconvenience (and sometimes physically harm) honest people and increase the likelihood of successful terrorism!
*Any* business based solely on patents is pure overhead for the rest of us. Overhead of the patent system, I mean. The fact that you can have a growth business of just dealing with patents and nothing else means that the overhead of the patent system itself grows over time. At some point the overhead costs more than the benefit that we get from the system! Don't get me started on insurance companies, lawyers, bankers and brokers either. Each system has players that are pure overhead, and increasingly so.
Just shoot raw and process the photos in Photoshop. Then their demosaic algorithm detector will just read "Adobe did it".
Finally we will be treated to live-action adaptations of such classics as Buck Rogers, Sherlock Holmes, Wonder Woman, Red Sonja, Barbarella, The Green Hornet and The Flash! Wait, what? Those were made into live-action years ago? Oh bloody hell. Oh well, I for one am looking forward to the sequels to recent films in that list, such as Iron Man and (maybe sequel) Superman. I guess I will treat the rest as not-really-sequels and / or pseudo-remakes. Lots of movies were better the second time around anyway. Most aren't, but here's hoping.
So child aggression before video games existed was caused by... what again? Oh, TV, right? And before that it was caused by... comics, right? And before that there was no child aggression.
The doctor would not proceed with the procedure until after we contacted our insurance carrier and got a letter of confirmation from them.
At least you didn't have a bu-reau-crat standing in between you and your doctor!
Will someone explain to me how a market can be truely free and not devolve into a conglomeration of companies screwing the populous out of their money through treatury and customer lock-in?
It can't. A free market will usually evolve into a cartel. The theory that a free market is good for everybody relies on the assumption that everyone involved has access to enough information to make informed decisions. This does not happen in real life, and in fact the cartels will try to hide the information that you need, unless you have some independent agency operating as you suggested. End of story.
I said it before and I'll say it again: Any model that is built on 10 or 20 positive examples from a population of 6,000,000,000 is going to suffer from overfitting. Not just a little overfitting... I mean it's going to overfit like a mo-fo. There's just no way, and I mean NO way, to create a statistically significant test based on the data we have on who is and who is not an ACTUAL terrorist.
Shadow Analysis Could Spot Known Individuals With Some Degree Of Accuracy.
Fixed it for you.
The new work IS nice, but it's not totally new.
Of course. Not much work is totally new. But it's new enough to be accepted into Siggraph, which is not an easy conference to get into.
NOT old news. Google for "2008 siggraph papers". Read the paper. Google for "2004 siggraph papers". Read about the old paper. Note the differences. Tim Rowley posts links to the papers from each year, so his site is recommended. Virtually all of these image-processing-related news items can be read long before they reach slashdot simply by keeping up with the latest papers from siggraph. In case you're lazy, the old paper is "Non-photorealistic Camera: Depth Edge Detection and Stylized Rendering Using a Multi-Flash Camera". Oddly, it's offline now. But I do have a copy of it on my hard drive. If you're not lazy, I HIGHLY recommend perusing all of the years' papers listed on Tim's site.
"The project's director thinks devices based on it could be implanted directly into the human body, in order to achieve super zoom vision!"
No, I would say the best way geeks can save the planet is to engineer a way to clean up the entire environment, as well as engineering a new energy infrastructure that needs less clean-up maintenance. Everything else is secondary to that.
But what do you patent? The thing you built? Someone else could build nearly the same thing without paying any royalties. The issue lies in how you define "nearly the same". Should it be the same on the quantum level? On a macroscopic level? The same components in the same relative configuration? In software, should it be the same machine code? The same source code? A logically equivalent algorithm? A similar user experience? This needs to be defined or you have nothing.
I called this when the PhysX cards first came out. I told my excited coworkers, "these cards are going to be irrelevant pretty soon, because it will all move to the GPU". They looked at me funny.
Hahaha, that's awesome. It's like the psychologists with the chimps choosing different colored M&Ms. Turns out sometimes you need to do math when you do statistics!
Any model that is built on 10 or 20 positive examples from a population of 6,000,000,000 is going to suffer from overfitting. Not just a little overfitting... I mean it's going to overfit like a mo-fo. There's just no way, and I mean NO way, to create a statistically significant test based on the data we have on who is and who is not an ACTUAL terrorist. Books like this are pure speculation.
And which test are you talking about, that gets more and more accurate the more times you apply it to an individual? Can you apply it to the same person 10000 times in one sitting, to get the required 99.9999% accuracy that it would need to be useful? Or do you need to invite them back to take the test on 10000 different occasions? Enlighten us more.
Seriously, guys. The article is from October 2005, you know, when the rest of the internet read it.
You should read the competition rules. The test set is so enormous that you would need 2^something_huge entries to see the results we've seen based on randomness. I did a back-of-the-envelope calculation at the beginning of the competition to see if a random search would be feasible to win the prize, and it's not. Not in a million years. Literally.
Wow, you sure told me. Straw man AND personal attacks... I guess you win.
If you can't build a large hadron supercollider in the game and get new insights into particle physics, in real time, then it fails the test. This is NOT near future.