This is demonstrably false. First of all, Scandinavian countries are overwhelmingly Lutheran. When you are born in Finland, Sweden, Denmark, or Norway, you are by default part of the Lutheran church. Even though they don't attend church regularly, the vast majority identify themselves as Christians and voluntarily choose to pay the optional church tax.
85% of Swedes are atheists (I'm one of them). And as for the Church tax, I'm paying it as well as it doesn't primarily go to the religious crap, but to the upkeep of the churches which are part of our historical heritage.
I argue that religion is fundamentally about morals. If you read the Jewish Bible, it's a code of law: what is right and what is wrong.
Which parts exactly? Like this one?
Six days shall work be done, but on the seventh day there shall be to you an holy day, a sabbath of rest to the LORD: whosoever doeth work therein shall be put to death. (Exd 35:2)
or perhaps
Your male and female slaves are to come from the nations around you; from them you may buy slaves. [Leviticus 25:44]
or
Samaria shall bear her guilt, because she has rebelled against her God; they shall fall by the sword, their little ones shall be dashed in pieces, and their pregnant women ripped open. [Hosea 13]
etc. Anybody getting their morals from that book is a lunatic.
And the obligatory reply is, who has been responsible for more mass murder? Christians or Atheists?
And who killed more people specifically because of their religious beliefs -- not political, paranoid, or power-hungry reasons -- Christians or Atheists?
The first question is flawed, the second is relevant. Atheists such as Stalin and Mao didn't kill people in the name of atheism. Many however have killed in the name of a religion.
Stalin had a moustache, so did Hitler. Who is responsible for more mass murder, people with moustaches or Christians?
Morality is fundamentally independent of religion - just look at the Scandinavian countries that are overwhealmingly atheist but have very little crime and violence. Compare that to the über-religious Middle East. In a normal state of mind people know that it is is wrong to murder, rape etc It takes strong faith to overcome that baseline morality, and religion can provide it. As physicist Steven Weinberg put it: "Normally good people do good things and bad people do bad things. It takes religion to make good people do bad things"
While it is a bit simplified, the statement stands - religion corrupts morality. While mainstream religious people choose to ignore the nasty parts of their religion they enable the extremists by advocating that faith (belief without evidence) is a good thing.
It is easy to get a neural net to perform with 100% accuracy on training data, because with enough parameters you can fit an elephant, and neural nets have oodles of parameters. But the reality is that the training data almost never fully encompasses the richness of the problem domain. Performance typically degrades badly on independent test data, and only very rarely stays as good as 90%. 75% is typical.
Um, not really. A text book example of neural nets is OCR where accuracy on a test set less than 99% is considered rubbish. For a bit more advanced methods > 99.9% accuracy on OCR is quite possible to achieve. And OCR is a 26-way classification rather than a simple binary one.
It is impossible to generalize about the accuracy in such a way as it depends on the distribution of the data and the number of samples as well as the difficulty of the problem. That goes for any statistical method, not just neural nets.
As for test vs validation data (the latter being a subset of the total data set and which is not used during training), in practice they are most likely to be the same. It is rare that you can make truly independent measurements to obtain a test set.
As far as I can see in the police example they are using a validation set that consists of a subset of the total data and that has not been shown to the neural net during training.
No, what is remarkable is the context of the data - predicting human behaviour with few relatively trivial parameters.
There are more subtle effects, too: because there are few good principles for network design, people often try different network architectures until they get one that performs well on the test data. In this case, even though they never formally train the net on the test data, they have effectively reduced the statistical confidence of the final result by re-using the test data until they get a good result on it.
Again, this depends on the distribution and quantity of data. If you capture a statistically significant amount of data that covers the whole input space distribution then re-using the test data isn't a problem.
By what we know of our own planet's past, it is pretty much universally accepted that humanity has not existed forever. So that creates the question of where we came from.
...and is quite well explained by evolution. For an overview our family tree, look at the Human evolution article on wikipedia.
The really nice thing about evolution is that unlike the designer hypothesis doesn't end with an infinite regress. The fundamental principle of natural selection is self-explanatory to such a degree that there is no room for a meta level. That organisms that are good at surviving are the ones that survive is really a principle difficult to dispute. And asking "Why do things that are good at surviving the ones that survive?" makes no sense. In essence the theory builds on such a simple axiom that it kills off the need for a meta level.
Physics isn't there yet - we don't have a solid scientific theory of the creation and development of the universe. Hopefully one day we'll have something as simple and elegant as Darwin's theory but right now we have a long way to go. Note however that just because we don't know how it works does not in any way support a supernatural explanation. Through a large part of the history of humanity we had no clue about how anything worked in the world. Now we know quite a bit more and not in one single instance has a supernatural explanation been the right one. So it would be silly to assume that the things we don't understand today will have a supernatural explanation.
I'm not following your reasoning. Why do you believe that there being a Creator explains nothing? There being a creator changes many things about the way we live life - the value of other humans, the nature of morality, their being an objective and external purpose to life.
There just being a creator tells us nothing. It doesn't say if it is a Sumerian creator or Greek creator or alien creator and hence it tells us nothing about morality or the "purpose to life". If you want to claim that it is a specific designer, like say the Flying Spaghetti Monster or the Christian God, then you have much more to prove than just that there is a designer. If you can prove that it is a specific creator with some known features then it might be more meaningful.
Besides, the question of who designed the designer is going to lead us nowhere. If we allow ourselves that question, then we will ask, "who designed the designer of the designer?" and so on. There are perfectly rational reasons for believing that the Creator is itself uncreated.
You are on the right track there, just take one step back and say that the question of who designed the world isn't going to lead as nowhere. If you are bent on not explaining anything then the more simple solution is to postulate that the universe has existed forever and is timeless. Why introduce the extra regress of a designer when it doesn't explain anything? And if you are hell-bent on having this meta-level why stop an obvious infinite regress at an arbitrary point? We can say as much about the hypothetical designer as we can about the hypothetical designer's hypothetical designer. So why just stop our questions at the level of the hypothetical designer?
There is a similar thing on this site showing how to predict which police cadets would become good cops and which would become bad cops. It's some form of neural net tutorial, but the conclusions are (at least to me) remarkable:
We can with 96.55% confidence say that a graduating cadet will be failing at his job in five years and we can with 99.17% confidence say that a graduating cadet will be performing an adequate job five years in the future.
Why should it be the responsibility of ID to explain who created the creator?
Because it otherwise fails to explain anything. If irreducibly complex things require a designer then the designer who designed them will be even more complex. Since the designer theory can't tell us, well, anything, the only way to investigate is to go up the ladder: who designed the designer?
If you say that that's a metaphysical question that cannot be answered, why not just skip the whole designer/creator bit and say that you are not interested in physical modeling of the world. Invoking an extremely improbable super-being to explain the world is very unhelpful. That's what earlier civilizations did: thunder was Thor riding in his carriage in the sky etc
What the ID followers want is a return to that using the logic "I don't understand it so it must be God's work."
Audiophiles have consistently been failing double blind tests when it comes to lossy vs lossless audio compression.
Now, if you wish to sell stuff to audiophiles, then players supporting lossless compression are excellent - they will buy it (along with anything you claim, on whatever grounds, will improve the playback quality).
If you however want to bring better music quality to the general population - make them get better headphones.
The article goes into a discussion about how the 'sanitized' information is so prevalent in Chinese education that it is seen as the 'truth'.
Wikipedia is based on the principle of "relevance by consensus". While there is a requirement for providing references, there is no mechanism for objectively accepting or rejecting a reference or a theory. This leads in many cases to fringe theories of some interest group getting more attention than they should. The english wikipedia has the benefit of being international so that the diversity is larger and hence the process of reaching consensus is more complicated.
In China those same principles yield different results as the Chinese consensus on many political issues is not the same as the western/international consensus.
To be fair this plagues mostly the social sciences. Politics is largely based on opinion (and you can find whatever references you like, there are plenty of them) and history is has always been subjectively inclusive.
At least as it looks right now, the only system that's even going to be an improvement over GPS is Galileo, and even then it won't be by much.
The key difference isn't technical, but a question of control.
The big thing with Galileo is that it won't be controlled by a military but by a civilian administration. Given the development of world affairs in recent years it's perfectly reasonable for the Europeans to want an independent system. Europe is very dependent on the current GPS system and the US has clearly shown that it can't be trusted in terms of international agreements. While it is unlikely that the US would disrupt the service for European users, it is not an impossibility. The results of such a disruption would be disastrous and it's quite understandable that the EU doesn't want to take that chance.
Here in Sweden the politicians are currently discussing a system where music and movie downloads would be legal while general tax would be introduced and the income distributed to the artists/record labels/movie studios. The big unresolved issue is how the incomes to the tax would be divided. You'd have to have statistics of the downloads which implies some form of centralized system, which is difficult in practice. Furthermore there are issues with international copyright laws etc. It is unlikely that this will be introduced in the near future and if it is introduced they will probably want to execute it at an EU level - which will take even more time.
There is however an immediate positive effect of these discussions - the politicians are beginning to understand that going after file sharers isn't an acceptable solution. The sheer mass of people that are sharing, and the very few lawsuits and convictions make the law basically arbitrary. And that's one thing that the law can't be - random.
In addition the trials so far have been spectacular failures for the music industry as it has been established that you can't get a search warrant on charges of copyright violation. This means that the police have no way of securing evidence against file sharers and the courts have established that IP logging and screen shots are insufficient evidence.
While it is far from over both legally and politically, both the legislative and judicial parts of the government are starting to look at the problem realistically.
So why aren't they used? The answer is two-fold. First of all Bayesian filters are very fast to train and very fast to use. Neural nets are computationally expensive to train and fast to use while support vector machines are expensive to both train and use.
The other reason is that apparently the people writing the mail clients have little or no knowledge of the more advanced methods while the people in the "AI" community seem to have limited interest in spam filtering.
Also, in the long term, server-side filtering is the only acceptable solution. Even with an adequate client-side spam filter, you have the problem that you are downloading the mail from the mail server. This not only puts unnecessary strain on the server but can be quite expensive if you for instance are synching your mail on your cellphone. And server-side anti-spam software is developed at an excruciatingly slow pace.
Finally, the second front must be legal. Wouldn't it be nice if the law enforcement agencies focused on getting the spammers rather than chasing file sharers? Unfortunately, there seems to be little interest for that in the US (the primary source of spam). In the EU it is illegal to send spam to somebody if you haven't gotten explicit permission from the person you are sending it to. In the US it isn't illegal unless the person you are sending it to hasn't explicitly forbidden you to do so. A change of the US system to the one they have in Europe would be preferable.
The Hungarian prime minister has been caught admitting to lying, which is a big difference from Bush who was just caught lying. A part of the social contract in a modern democracy is for those in power to pretend that they are always doing what they think is right and for the voters to pretend that they believe it. A one-sided breach, like the voters thinking they have been lied to, can be handled, but when both sides break it, well, you get riots.
This is not legally possible. You cannot sign away your privacy protections, just like you cannot sign away workplace protection rules for example.
You wouldn't be signing away the privacy protection, you would be giving away the actual data. Regardless of the protection laws, I still have the right to give away my personal information to whom I see fit.
Indeed. The US has the right to decide on which terms people can enter, regardless of how absurd or not they are. And the EU has the right set privacy laws for EU companies which of course they must follow.
The solution is pretty obvious - the passengers have to sign an agreement granting the airline they are flying with to give out the personal information without any guarantees that the information will be protected.
However, while in case of tourists the passengers have the choice to tell the US and the airlines to go screw themselves - and opt not to travel to the US, in many cases this won't be a real option. If it is a business trip then they have no choice in practice. This opens up a whole new can of worms (on the EU side) in terms of a company de facto forcing its employees to turn over private information. This would be illegal as well in the EU, making it difficult for a company to send an employee on a business trip to the US.
To my non-US friends. The RIAA and MPAA are non-governmental, private industry groups. They have nothing to do with the US government, nor do they take direction from the US government.
No, it is the other way around - the government takes direction from them. And that's the problem. Rather than being driven by the priorities of the people the government's decision making is driven by special interest groups.
In Sweden we have had two legal cases regarding file sharing. The first one what a guy who shared a Swedish made movie on a p2p network. The verdict was that it was illegal, but only to such extent that it warranted a relatively small fine. By setting this limit, the courts defined in accordance to Swedish law that in future cases regarding file sharing no search warrants would be issued - as the crime wasn't serious enough. In addition they said that collecting data (such as IP-addresses) in order to prove a piracy case was in violation of Swedish privacy laws. So in effect they said "It's not OK in theory, but it's not worth using police and other resources for".
The second case was the now infamous pirate bay case. This was a direct government intervention by the ministry of justice as a result of US government pressure (on behalf of RIAA of course). This backfired badly as people here mostly think of lobbying as a form of corruption. And when government ministers got involved in ordering police around, it was even more than that - it was a blatant violation of the constitution. We have elections this year, and most likely in part due to this incident we'll see a "regime change". In the fall, the first hearings in this case will start for the impeachment of those that probably violated the law by ordering those raids on the pirate bay.
The point of this story being that what ultimately matters is how willing the government is willing to listen to the RIAA vs. listening to the people (or common sense) when deciding if piracy is a serious enough crime to deserve resources being allocated for fighting it.
What Google is essentially doing is how parents get their kids to do house chores: "Let's play a game: You clean the kitchen and your brother cleans the living room and we'll see who's faster! After that we can play 'Mowing the Lawn Game'!"
What they apparently plan to do is to use the data to train some supervised adaptive system, like a neural network to perform image classification. This type of data collection is usually very tedious (or expensive if you hire somebody else to do it). And google, the clever "parent" has managed to get its "children" to willingly work.
No you can't. You need the variance to compute significance levels, which we don't have. Either way we don't need it as we can't reject the null hypothesis anyway due to the too small number of samples.
It most certainly is not. If the probability of spontaneous remission were zero then it'd be impossible to get 2 out of 17 spontaneous remissions so we'd be 100% sure that the two remissions were not spontaneous.
No, I didn't say that the probability of spontaneous remission would be zero, but that the probability for the medicine could be it, given the margin of error. That the positive hits had some completely different cause and were only included because of a too small sample size - which in turn resulted that by chance a far more improbable cause dominated the results of the sample.
You're saying something completely bizarre. After you've collected 1000 sample posts you'll either catch my posts or not. If you do, then you'll be able to reject the hypothesis as clearly false.
With mathematical logic, yes, with statistics, no. With 1 out of 1000 samples being true we would with 99% certainty not be able to reject the hypothesis (that no such signatures exist).
c = +-sqrt(2.58^2 * (1/1000)*(1-1/1000)/1000) = 2.6/1000. With one positive sample the confidence interval would be 1+-2.6 of 1000. This encompasses zero probability (e.g the null hypothesis) hence we can't reject it at a 99% confidence level.
Had we found 2 posts of 1000, then we would still confirm the hypothesis at a 95% level (z=1.96), but we would reject it at a 80% level (z=1.28)
You have to understand that confirming or rejecting a hypothesis at a confidence level (statistics) has very little to do with confirming or rejecting a hypothesis using deductive logic.
Statistics deals with probabilities and is never absolute. You don't say that a hypothesis is true or false, but that it is true or false at a certain confidence level. And that's what's great about it as real-world science deals with real-world measurements which are always associated with various forms of errors and weird correlations that you don't want influencing your results. Sampling to obtain a probability estimate is a very convenient tool, but if you take too few samples, the margin of error will be to great, and if it encompasses zero, then it is worthless. An estimate that doesn't reject the null hypothesis means that the estimate is no better than a random guess.
We don't have a particularly good estimate of the rate of remission caused by the new therapy.
Exactly, zero rate of remission is within the margin of error. By that definition the null hypothesis can't be rejected. Get it? No matter how small the spontaneous remission rates are, they will be higher than zero. I'm not sure how to explain it to you further - I've tried both with words and equations.
No. But (1) I think you need the some skill in the former before you can embark on the latter and (2) in a suitable limit, statistics needs to match mathematical logic, and a good test of your statistical methodology is to examine it's behaviour in the limit and compare it with mathematical logic.
Um, no, they operate at completely different levels. To give you an example, let's suppose that our hypothesis is that there are no internet users that have "Reinstate Pluto!" in their signature when posting in forums. Using a logical proof it's simple - we must reject the hypothesis. You have it in your sig so clearly the hypothesis is false. The statistical conclusion will be the exact opposite. After we sample 1000 random forum posts across the Internet, we will be able to confirm the hypothesis with 99.9% confidence or more.
Good enough to publish
You'd be surprised how ignorant medical researchers are of statistics and what complete rubbish gets peer-reviewed and published. I work for a company that does data mining and we've done consulting on way too many medical research projects for me to have any illusion about how the scientific method is applied in that field. I can't tell you how many times I've had to explain to doctors that no, you can't use two samples for building your model and a third for validation. I do sympathize though, there's no tougher field of data collection both because the human factor and the fact that truly random samples are very difficult to obtain.
But we have excellent evidence that it is indeed causing remissions.
No we don't. The actual spontaneous remission rate for this type of cancer (malignant melanoma) is 1/400. I looked it up here. Now if we do a little thought experiment. Suppose we took 10 groups of people with 40 people in each. We give each group a different type of soft drink. After a while we notice that of those drinking fanta, we had one survivor. We are terribly pleased with ourselves as obviously we have found that fanta cures cancer in 1/40 people compared to the 1/400 which is the standard going rate. Or have we?
No, probably fanta didn't cure the cancer. Most likely it was just a case of being within the error margin of the sampling. Or would you say in this case that we have "excellent evidence that it is indeed causing remissions"? The overall probability for the population is one thing, but the confidence of a measurement of that probability on your sample is a different thing. The larger the sample size the closer you will get to the real value. Calculating the error margin gives you how close you are to the probability for the population. In this example your estimate may have nothing to do with the real probability, as zero probability is within the margin of error. So in fact, since the spontaneous rate is > 0, for all that you know your medicine may actually decrease the chance of survival for your patients.
It's quite simple to calculate the sample size needed to come outside the margin of error. I gave you the equation. If you wish the confidence to be 2+-1 of 17, which is the minimum then you need 278 samples. Even if you would assume that the total number of possible cases to pick your 17 from was as low as 100, you'd need 74 samples to say with 95% confidence that 2+-1 of 17 were cured.
As for your example, you are wrong because you are confusing mathematical logic and statistics. Claiming that something impossible you need only one example to prove otherwise, while to prove that something with probability lim->0 p you need N number of examples where lim N->infinity. And again, you are confusing significance level with error margin. The former is dependent on prior probabilities while the latter is not. For calculating the significance level, the rate of spontaneous remissions would be the key, but for the margin of error it makes no difference. The margin of error is due to the inevitability of sampling error which applies to any random sampling of a finite or infinite population.
Getting above the sampling error is the minimum requirement, needed before you start discussing significance levels.
That's statistical significance. The calculation that I showed you is for sampling error, and that only depends on the sample size and the measured posterior probabilities of the outcome (population size can be taken into consideration for finite populations).
Since 2 of 17 people survived, that's your positive outcome probability. For sampling error the likelihood of of spontaneous recovery makes no difference at all. If we wanted to calculate a significance level, it would - but in this case it isn't necessary as the sampling error is too high anyway.
Simply put, their sample size was way too small.
For an equivalent situation, imagine you ask 17 random people on the on which of two political parties they would vote. If two said party A and 15 said party B, would you think that it would be a good idea to publish that as a reliable result? It's not a coincidence that polling organizations use sample sizes of 1000+ people when they gather statistics. This is no different - to avoid being within the margin of error, the researchers should have used far more samples. It would have not assured statistical significance, but they would have passed the first obstacle - the margin of error. As they didn't even bother with this, I would be surprised if they did any estimates of the significance level.
Just a correction, what I meant to say was that it isn't a normal distribution per se (as there are are only two possible outcomes: dead or not dead). This is however exactly the same calculation you use for instance with polling when you calculate an error margin. The only information you have available is sample size and the probabilities of the different outcomes.
If you don't trust my equations, you have them here and on the same site you can find a javascript sample size calculator which you can test the values with.
As several people have posted, the chance of spontaneous remission is about half of this result. As for the confidence calculation, it is fairly simple and of course it is a normal distribution.
The simple calculation is as follows:
Sample size: ss=17
Positive outcome probability: p = 2/17
Sample size = ss = Z^2(p)*(1-p)/c^2
Z = z value (e.g 1.96 for 95% confidence)
c = confidence interval for probability
so
c = +- sqrt(Z^2*(p)*(1-p)/ss)
Insert values: c = +-sqrt( 1.96^2 * (2/17)*(1-2/17)/17) = +-0.1532
So the probability is 2/17 +- 0.1532 with 95% confidence or expressed for a sample of 17:
Slashdot Mirror
The first question is flawed, the second is relevant. Atheists such as Stalin and Mao didn't kill people in the name of atheism. Many however have killed in the name of a religion.
Stalin had a moustache, so did Hitler. Who is responsible for more mass murder, people with moustaches or Christians?
Morality is fundamentally independent of religion - just look at the Scandinavian countries that are overwhealmingly atheist but have very little crime and violence. Compare that to the über-religious Middle East. In a normal state of mind people know that it is is wrong to murder, rape etc It takes strong faith to overcome that baseline morality, and religion can provide it. As physicist Steven Weinberg put it: "Normally good people do good things and bad people do bad things. It takes religion to make good people do bad things"
While it is a bit simplified, the statement stands - religion corrupts morality. While mainstream religious people choose to ignore the nasty parts of their religion they enable the extremists by advocating that faith (belief without evidence) is a good thing.
It is impossible to generalize about the accuracy in such a way as it depends on the distribution of the data and the number of samples as well as the difficulty of the problem. That goes for any statistical method, not just neural nets.
As for test vs validation data (the latter being a subset of the total data set and which is not used during training), in practice they are most likely to be the same. It is rare that you can make truly independent measurements to obtain a test set.
As far as I can see in the police example they are using a validation set that consists of a subset of the total data and that has not been shown to the neural net during training.
No, what is remarkable is the context of the data - predicting human behaviour with few relatively trivial parameters.
Again, this depends on the distribution and quantity of data. If you capture a statistically significant amount of data that covers the whole input space distribution then re-using the test data isn't a problem.
The really nice thing about evolution is that unlike the designer hypothesis doesn't end with an infinite regress. The fundamental principle of natural selection is self-explanatory to such a degree that there is no room for a meta level. That organisms that are good at surviving are the ones that survive is really a principle difficult to dispute. And asking "Why do things that are good at surviving the ones that survive?" makes no sense. In essence the theory builds on such a simple axiom that it kills off the need for a meta level.
Physics isn't there yet - we don't have a solid scientific theory of the creation and development of the universe. Hopefully one day we'll have something as simple and elegant as Darwin's theory but right now we have a long way to go. Note however that just because we don't know how it works does not in any way support a supernatural explanation. Through a large part of the history of humanity we had no clue about how anything worked in the world. Now we know quite a bit more and not in one single instance has a supernatural explanation been the right one. So it would be silly to assume that the things we don't understand today will have a supernatural explanation.
There just being a creator tells us nothing. It doesn't say if it is a Sumerian creator or Greek creator or alien creator and hence it tells us nothing about morality or the "purpose to life". If you want to claim that it is a specific designer, like say the Flying Spaghetti Monster or the Christian God, then you have much more to prove than just that there is a designer. If you can prove that it is a specific creator with some known features then it might be more meaningful.
You are on the right track there, just take one step back and say that the question of who designed the world isn't going to lead as nowhere. If you are bent on not explaining anything then the more simple solution is to postulate that the universe has existed forever and is timeless. Why introduce the extra regress of a designer when it doesn't explain anything? And if you are hell-bent on having this meta-level why stop an obvious infinite regress at an arbitrary point? We can say as much about the hypothetical designer as we can about the hypothetical designer's hypothetical designer. So why just stop our questions at the level of the hypothetical designer?
If you say that that's a metaphysical question that cannot be answered, why not just skip the whole designer/creator bit and say that you are not interested in physical modeling of the world. Invoking an extremely improbable super-being to explain the world is very unhelpful. That's what earlier civilizations did: thunder was Thor riding in his carriage in the sky etc
What the ID followers want is a return to that using the logic "I don't understand it so it must be God's work."
Now, if you wish to sell stuff to audiophiles, then players supporting lossless compression are excellent - they will buy it (along with anything you claim, on whatever grounds, will improve the playback quality).
If you however want to bring better music quality to the general population - make them get better headphones.
Wikipedia is based on the principle of "relevance by consensus". While there is a requirement for providing references, there is no mechanism for objectively accepting or rejecting a reference or a theory. This leads in many cases to fringe theories of some interest group getting more attention than they should. The english wikipedia has the benefit of being international so that the diversity is larger and hence the process of reaching consensus is more complicated.
In China those same principles yield different results as the Chinese consensus on many political issues is not the same as the western/international consensus.
To be fair this plagues mostly the social sciences. Politics is largely based on opinion (and you can find whatever references you like, there are plenty of them) and history is has always been subjectively inclusive.
The key difference isn't technical, but a question of control.
The big thing with Galileo is that it won't be controlled by a military but by a civilian administration. Given the development of world affairs in recent years it's perfectly reasonable for the Europeans to want an independent system. Europe is very dependent on the current GPS system and the US has clearly shown that it can't be trusted in terms of international agreements. While it is unlikely that the US would disrupt the service for European users, it is not an impossibility. The results of such a disruption would be disastrous and it's quite understandable that the EU doesn't want to take that chance.
There is however an immediate positive effect of these discussions - the politicians are beginning to understand that going after file sharers isn't an acceptable solution. The sheer mass of people that are sharing, and the very few lawsuits and convictions make the law basically arbitrary. And that's one thing that the law can't be - random.
In addition the trials so far have been spectacular failures for the music industry as it has been established that you can't get a search warrant on charges of copyright violation. This means that the police have no way of securing evidence against file sharers and the courts have established that IP logging and screen shots are insufficient evidence.
While it is far from over both legally and politically, both the legislative and judicial parts of the government are starting to look at the problem realistically.
So why aren't they used? The answer is two-fold. First of all Bayesian filters are very fast to train and very fast to use. Neural nets are computationally expensive to train and fast to use while support vector machines are expensive to both train and use.
The other reason is that apparently the people writing the mail clients have little or no knowledge of the more advanced methods while the people in the "AI" community seem to have limited interest in spam filtering.
Also, in the long term, server-side filtering is the only acceptable solution. Even with an adequate client-side spam filter, you have the problem that you are downloading the mail from the mail server. This not only puts unnecessary strain on the server but can be quite expensive if you for instance are synching your mail on your cellphone. And server-side anti-spam software is developed at an excruciatingly slow pace.
Finally, the second front must be legal. Wouldn't it be nice if the law enforcement agencies focused on getting the spammers rather than chasing file sharers? Unfortunately, there seems to be little interest for that in the US (the primary source of spam). In the EU it is illegal to send spam to somebody if you haven't gotten explicit permission from the person you are sending it to. In the US it isn't illegal unless the person you are sending it to hasn't explicitly forbidden you to do so. A change of the US system to the one they have in Europe would be preferable.
..no wireless, no 3G, no bluetooth, no GPS and 24 mb memory. What year do they think this is? 1987?
The Hungarian prime minister has been caught admitting to lying, which is a big difference from Bush who was just caught lying. A part of the social contract in a modern democracy is for those in power to pretend that they are always doing what they think is right and for the voters to pretend that they believe it. A one-sided breach, like the voters thinking they have been lied to, can be handled, but when both sides break it, well, you get riots.
You wouldn't be signing away the privacy protection, you would be giving away the actual data. Regardless of the protection laws, I still have the right to give away my personal information to whom I see fit.
The solution is pretty obvious - the passengers have to sign an agreement granting the airline they are flying with to give out the personal information without any guarantees that the information will be protected.
However, while in case of tourists the passengers have the choice to tell the US and the airlines to go screw themselves - and opt not to travel to the US, in many cases this won't be a real option. If it is a business trip then they have no choice in practice. This opens up a whole new can of worms (on the EU side) in terms of a company de facto forcing its employees to turn over private information. This would be illegal as well in the EU, making it difficult for a company to send an employee on a business trip to the US.
In Sweden we have had two legal cases regarding file sharing. The first one what a guy who shared a Swedish made movie on a p2p network. The verdict was that it was illegal, but only to such extent that it warranted a relatively small fine. By setting this limit, the courts defined in accordance to Swedish law that in future cases regarding file sharing no search warrants would be issued - as the crime wasn't serious enough. In addition they said that collecting data (such as IP-addresses) in order to prove a piracy case was in violation of Swedish privacy laws. So in effect they said "It's not OK in theory, but it's not worth using police and other resources for".
The second case was the now infamous pirate bay case. This was a direct government intervention by the ministry of justice as a result of US government pressure (on behalf of RIAA of course). This backfired badly as people here mostly think of lobbying as a form of corruption. And when government ministers got involved in ordering police around, it was even more than that - it was a blatant violation of the constitution. We have elections this year, and most likely in part due to this incident we'll see a "regime change". In the fall, the first hearings in this case will start for the impeachment of those that probably violated the law by ordering those raids on the pirate bay.
The point of this story being that what ultimately matters is how willing the government is willing to listen to the RIAA vs. listening to the people (or common sense) when deciding if piracy is a serious enough crime to deserve resources being allocated for fighting it.
Really sad. Australia has lost one of its greatest cultural icons, only comparable to Yahoo Serious or Paul "Crocodile Dundee" Hogan.
What they apparently plan to do is to use the data to train some supervised adaptive system, like a neural network to perform image classification. This type of data collection is usually very tedious (or expensive if you hire somebody else to do it). And google, the clever "parent" has managed to get its "children" to willingly work.
I'd demand payment in google stocks.
With mathematical logic, yes, with statistics, no. With 1 out of 1000 samples being true we would with 99% certainty not be able to reject the hypothesis (that no such signatures exist). c = +-sqrt(2.58^2 * (1/1000)*(1-1/1000)/1000) = 2.6/1000. With one positive sample the confidence interval would be 1+-2.6 of 1000. This encompasses zero probability (e.g the null hypothesis) hence we can't reject it at a 99% confidence level.
Had we found 2 posts of 1000, then we would still confirm the hypothesis at a 95% level (z=1.96), but we would reject it at a 80% level (z=1.28) You have to understand that confirming or rejecting a hypothesis at a confidence level (statistics) has very little to do with confirming or rejecting a hypothesis using deductive logic.
Statistics deals with probabilities and is never absolute. You don't say that a hypothesis is true or false, but that it is true or false at a certain confidence level. And that's what's great about it as real-world science deals with real-world measurements which are always associated with various forms of errors and weird correlations that you don't want influencing your results. Sampling to obtain a probability estimate is a very convenient tool, but if you take too few samples, the margin of error will be to great, and if it encompasses zero, then it is worthless. An estimate that doesn't reject the null hypothesis means that the estimate is no better than a random guess.
Exactly, zero rate of remission is within the margin of error. By that definition the null hypothesis can't be rejected. Get it? No matter how small the spontaneous remission rates are, they will be higher than zero. I'm not sure how to explain it to you further - I've tried both with words and equations.
Um, no, they operate at completely different levels. To give you an example, let's suppose that our hypothesis is that there are no internet users that have "Reinstate Pluto!" in their signature when posting in forums. Using a logical proof it's simple - we must reject the hypothesis. You have it in your sig so clearly the hypothesis is false. The statistical conclusion will be the exact opposite. After we sample 1000 random forum posts across the Internet, we will be able to confirm the hypothesis with 99.9% confidence or more. You'd be surprised how ignorant medical researchers are of statistics and what complete rubbish gets peer-reviewed and published. I work for a company that does data mining and we've done consulting on way too many medical research projects for me to have any illusion about how the scientific method is applied in that field. I can't tell you how many times I've had to explain to doctors that no, you can't use two samples for building your model and a third for validation. I do sympathize though, there's no tougher field of data collection both because the human factor and the fact that truly random samples are very difficult to obtain. No we don't. The actual spontaneous remission rate for this type of cancer (malignant melanoma) is 1/400. I looked it up here. Now if we do a little thought experiment. Suppose we took 10 groups of people with 40 people in each. We give each group a different type of soft drink. After a while we notice that of those drinking fanta, we had one survivor. We are terribly pleased with ourselves as obviously we have found that fanta cures cancer in 1/40 people compared to the 1/400 which is the standard going rate. Or have we?No, probably fanta didn't cure the cancer. Most likely it was just a case of being within the error margin of the sampling. Or would you say in this case that we have "excellent evidence that it is indeed causing remissions"? The overall probability for the population is one thing, but the confidence of a measurement of that probability on your sample is a different thing. The larger the sample size the closer you will get to the real value. Calculating the error margin gives you how close you are to the probability for the population. In this example your estimate may have nothing to do with the real probability, as zero probability is within the margin of error. So in fact, since the spontaneous rate is > 0, for all that you know your medicine may actually decrease the chance of survival for your patients.
As for your example, you are wrong because you are confusing mathematical logic and statistics. Claiming that something impossible you need only one example to prove otherwise, while to prove that something with probability lim->0 p you need N number of examples where lim N->infinity. And again, you are confusing significance level with error margin. The former is dependent on prior probabilities while the latter is not. For calculating the significance level, the rate of spontaneous remissions would be the key, but for the margin of error it makes no difference. The margin of error is due to the inevitability of sampling error which applies to any random sampling of a finite or infinite population.
Getting above the sampling error is the minimum requirement, needed before you start discussing significance levels.
Since 2 of 17 people survived, that's your positive outcome probability. For sampling error the likelihood of of spontaneous recovery makes no difference at all. If we wanted to calculate a significance level, it would - but in this case it isn't necessary as the sampling error is too high anyway.
Simply put, their sample size was way too small.
For an equivalent situation, imagine you ask 17 random people on the on which of two political parties they would vote. If two said party A and 15 said party B, would you think that it would be a good idea to publish that as a reliable result? It's not a coincidence that polling organizations use sample sizes of 1000+ people when they gather statistics. This is no different - to avoid being within the margin of error, the researchers should have used far more samples. It would have not assured statistical significance, but they would have passed the first obstacle - the margin of error. As they didn't even bother with this, I would be surprised if they did any estimates of the significance level.
If you don't trust my equations, you have them here and on the same site you can find a javascript sample size calculator which you can test the values with.
The simple calculation is as follows: Sample size: ss=17 Positive outcome probability: p = 2/17
Sample size = ss = Z^2(p)*(1-p)/c^2
Z = z value (e.g 1.96 for 95% confidence)
c = confidence interval for probability so
c = +- sqrt(Z^2*(p)*(1-p)/ss)
Insert values: c = +-sqrt( 1.96^2 * (2/17)*(1-2/17)/17) = +-0.1532
So the probability is 2/17 +- 0.1532 with 95% confidence or expressed for a sample of 17:
2 +-0.1532*17 of 17 = 2 +- 2.6 of 17
Elementary my dear Watson.