Exactly. And these systems are often tested on databases which have a large variety of faces photographed under identical conditions to get their headline 95% accuracy figure.
I did an experiment with a database of only 20 similar looking people under a range of conditions with humans doing the recognition and the best result was just over 80%.
New Scientist covered this in an article several weeks ago. The experiment is about retrocausality - the future affecting the past - rather than sending information back in time.
A few weeks later they also printed a letter which suggested retrocausality was redundant.
I don't pretend to understand this stuff. The message I got from the article though was that the future affecting the past only seems weird because we assume that time flows, because that's how we perceive it: if you imagine all time existing like a spatial dimension, then retrocausality is "normal".
Hey, sounds interesting. I implemented something similar when I was working on handwriting systems and wanted to send "handwritten emails". The email was just a notification with a URL to the handwritten page.
It worked pretty well. One big plus was that the sender could tell if and when the message had been read! (or at least viewed)
Obviously each sender had to have access to a server, which is a downer in some cases but should be fine for a web-based system.
Actually, it's not quite that bad. Obviously if you know some classifiers are less than 50% then you should ignore them; but you might only lose a bit of performance if the odd outlier slips in.
For example, two 80% classifiers and a 30% classifier would give a majority performance of about 74%
(one 80% classifier and two 30% classifiers will only get you 43%)
And obviously (I hope) we are talking about the true accuracy here, not the sampled accuracy over some small number of trials!
I'm not sure what you mean by aggregating probabilities, but I think you can only combine weak classifiers and get better results if all the classifiers have greater than 50% accuracy.
Say I have 3 classifiers which are 51% accurate. If I take their predictions as votes and go with the majority then that will give me an accuracy of 51.5% (0.51 * 0.51 * 0.49 * 3 + 0.51 * 0.51 * 0.51)
A majority of 101 classifiers would get you 58% and 1001 about 74%. The more classifiers the closer to 100% you could get.
But if you have some classifiers with less than 50% accuracy then the best you can do is ignore them!
Thanks for not advising me. But you assume too much. I didn't say I don't have a job and life insurance. I even play the lottery from time to time...
There's plenty of time in a day to earn money and daydream too!
Yes, very long copyright is a bad thing (I think 50 years in total is easily enough) but it is not rational to dismiss an incentive which only appeals to irrational people, if you know the people you want to incentivise are irrational.
...and everyone I know is prone to being irrational.
Because often an artist does not make much (or any) money from his/her work until years after it was done.
I'm not suggesting artists should be paid more than anyone else; just that they (or their children) should be paid.
Yes, I agree, a fixed term (50 years) would probably be better. But maybe that might encourage well known artists to stockpile work for their kids to publish?
"To believe that anything beyond lifetime copyright impacts the incentive structure of creators takes some serious suspension of disbelief."
I'm guessing you are quite young then? I'm over 40 and have 2 children. I write but haven't made a penny at it. Much as I enjoy writing, there is definitely an incentive to "publish" in knowing that if my work becomes popular late in my life (or just after) then my kids will benefit.
Can non-US numbers be registered? I am registered with the UK version (TPS) and don't get calls from UK telemarketers, but I do get calls from US telemarketers... and there's nothing the TPS can do about it.
Mostly I get calls in the early evening; which is really annoying.
If you look at an S-curve before the inflexion point, you can convince yourself that it's going to keep on rising and become exponential. That's the problem with extrapolation... it's bollocks (or not) you don't know, you can't know.
I think you are right about Yahoo wanting to sell to iPod owners. In which case, maybe Apple have done everyone a favour by not licensing their DRM... they may kill DRM!
I wonder if that was their plan all along? Apple make a lot more money from selling iPods than they do from selling iTunes, don't they? And people would still use iTunes if it sold MP3s because it integrates well with their iPod... in fact, I might start using it if it sold MP3s.
A patent doesn't mean you own an idea (which I agree would be absurd). A patent gives you a monopoly on the commercial exploitation of an idea for a limited time.
Also, if I'm already using an idea (but have kept it secret) and you come along and have the same idea and patent it, I can still use it (as long as I can prove I was using it before you filed your patent).
Patents don't deal with any specific representation, copyrights do.
A patent is an abstract generalisation of an idea, which must have a "technical effect". You can't patent the code "0101111000" but you can patent a "method of counting pips in an apple". You only have to describe an actual implementation as an illustration; the guts of the patent are the "claims" which are implementation-independent.
This is how lawyers patent software in the UK. You patent a "machine" which performs a sequence of steps. Then you can argue later that software running on a computer is an embodiment of that machine and infringes your patent.
I am not a lawyer. I have worked with a few... and helped to patent software. [I'm sorry]
How about: lighting changes (sun comes out / goes in), shadows cast by passing objects, reflections from moving objects, camera auto-gain triggered by scene composition changes, camera noise, white-out.
Big problem. You can make it work some of the time...
Indeed. You might think "why not just pick another name? how hard can it be?" as I did until recently when I came across a company at a trade show with a new software product... with no name.
The lawyers had chucked out all their suggestions. So there was a product (almost) ready; and no name. I heard a reseller say "look, I need a name if I'm gonna sell this!"
Now I know why the next big thing is the XKI-231047
The New Scientist had a more detailed article on this last month. It also has a link to an AVI of a bee flying and a reference to the full journal paper.
Not again! I worked on face recognition for several years, for applications unrelated to security (e.g. searching photo collections for family members). Time and time again people said "Hey, you could use this for access control!" and would't listen when I pointed out that you would be lucky to get a recognition accuracy of 70% in real-world conditions.
I've implemented methods which claimed a 99.X% recognition rate and found the real-world results were often as low as 60%... I assume people don't lie when they publish these things, but they clearly construct their test sets very carefully:-)
Sure, you can make a system which stops a blonde woman from accessing a dark-haired man's phone; but distinguishing between two similar looking people and still allowing an individual's apperance to vary is not currently possible (even for a lot of humans!)
I work for a software company, most of the people I work with are software engineers. They are all very professional - they all seem happy and the company seems happy with them.
We have an IT department that looks after systems administration - I've only met one of them, once... they are very efficient but hide somewhere and only accept questions via phone or a web-form.
I have friends who call themselves IT consultants - they are so professional that they don't even have to write any software, ever... they make presentations about it and talk a lot. They don't care about being valued because they get paid loads.
Slashdot Mirror
Exactly. And these systems are often tested on databases which have a large variety of faces photographed under identical conditions to get their headline 95% accuracy figure.
I did an experiment with a database of only 20 similar looking people under a range of conditions with humans doing the recognition and the best result was just over 80%.
New Scientist covered this in an article several weeks ago. The experiment is about retrocausality - the future affecting the past - rather than sending information back in time.
A few weeks later they also printed a letter which suggested retrocausality was redundant.
I don't pretend to understand this stuff. The message I got from the article though was that the future affecting the past only seems weird because we assume that time flows, because that's how we perceive it: if you imagine all time existing like a spatial dimension, then retrocausality is "normal".
There are some Japanese phones like this for example
I guess they are just way ahead of "us".
Hey, sounds interesting. I implemented something similar when I was working on handwriting systems and wanted to send "handwritten emails". The email was just a notification with a URL to the handwritten page.
It worked pretty well. One big plus was that the sender could tell if and when the message had been read! (or at least viewed)
Obviously each sender had to have access to a server, which is a downer in some cases but should be fine for a web-based system.
Best of luck with your project.
Actually, it's not quite that bad. Obviously if you know some classifiers are less than 50% then you should ignore them; but you might only lose a bit of performance if the odd outlier slips in.
For example, two 80% classifiers and a 30% classifier would give a majority performance of about 74%
(one 80% classifier and two 30% classifiers will only get you 43%)
And obviously (I hope) we are talking about the true accuracy here, not the sampled accuracy over some small number of trials!
I'm not sure what you mean by aggregating probabilities, but I think you can only combine weak classifiers and get better results if all the classifiers have greater than 50% accuracy.
Say I have 3 classifiers which are 51% accurate. If I take their predictions as votes and go with the majority then that will give me an accuracy of 51.5% (0.51 * 0.51 * 0.49 * 3 + 0.51 * 0.51 * 0.51)
A majority of 101 classifiers would get you 58% and 1001 about 74%. The more classifiers the closer to 100% you could get.
But if you have some classifiers with less than 50% accuracy then the best you can do is ignore them!
Thanks for not advising me. But you assume too much. I didn't say I don't have a job and life insurance. I even play the lottery from time to time...
There's plenty of time in a day to earn money and daydream too!
Yes, very long copyright is a bad thing (I think 50 years in total is easily enough) but it is not rational to dismiss an incentive which only appeals to irrational people, if you know the people you want to incentivise are irrational.
Because often an artist does not make much (or any) money from his/her work until years after it was done.
I'm not suggesting artists should be paid more than anyone else; just that they (or their children) should be paid.
Yes, I agree, a fixed term (50 years) would probably be better. But maybe that might encourage well known artists to stockpile work for their kids to publish?
I'm guessing you are quite young then? I'm over 40 and have 2 children. I write but haven't made a penny at it. Much as I enjoy writing, there is definitely an incentive to "publish" in knowing that if my work becomes popular late in my life (or just after) then my kids will benefit.
Parents want to provide for their children.
Mostly I get calls in the early evening; which is really annoying.
That's even if you're measuring the right thing.
I wonder if that was their plan all along? Apple make a lot more money from selling iPods than they do from selling iTunes, don't they? And people would still use iTunes if it sold MP3s because it integrates well with their iPod... in fact, I might start using it if it sold MP3s.
Also, if I'm already using an idea (but have kept it secret) and you come along and have the same idea and patent it, I can still use it (as long as I can prove I was using it before you filed your patent).
I recall this happened in two seats in this year's local elections (where the number of electors is small and the turnout is even smaller).
One of the losers was interviewed on the radio and seemed quite cool about it - "fair enough", he said.
A patent is an abstract generalisation of an idea, which must have a "technical effect". You can't patent the code "0101111000" but you can patent a "method of counting pips in an apple". You only have to describe an actual implementation as an illustration; the guts of the patent are the "claims" which are implementation-independent.
This is how lawyers patent software in the UK. You patent a "machine" which performs a sequence of steps. Then you can argue later that software running on a computer is an embodiment of that machine and infringes your patent.
I am not a lawyer. I have worked with a few... and helped to patent software. [I'm sorry]
How about: lighting changes (sun comes out / goes in), shadows cast by passing objects, reflections from moving objects, camera auto-gain triggered by scene composition changes, camera noise, white-out.
Big problem. You can make it work some of the time...
The lawyers had chucked out all their suggestions. So there was a product (almost) ready; and no name. I heard a reseller say "look, I need a name if I'm gonna sell this!"
Now I know why the next big thing is the XKI-231047
You think the earth isn't flat?
You think space is flat?
You think time is what?
It's just a theory? OK, call me when you've proved it...
The New Scientist had a more detailed article on this last month. It also has a link to an AVI of a bee flying and a reference to the full journal paper.
Not even nearly funny.
You can leave out one of xsize or ysize if you want to keep the aspect ratio the same...
Not again! I worked on face recognition for several years, for applications unrelated to security (e.g. searching photo collections for family members). Time and time again people said "Hey, you could use this for access control!" and would't listen when I pointed out that you would be lucky to get a recognition accuracy of 70% in real-world conditions. I've implemented methods which claimed a 99.X% recognition rate and found the real-world results were often as low as 60%... I assume people don't lie when they publish these things, but they clearly construct their test sets very carefully :-)
Sure, you can make a system which stops a blonde woman from accessing a dark-haired man's phone; but distinguishing between two similar looking people and still allowing an individual's apperance to vary is not currently possible (even for a lot of humans!)
I work for a software company, most of the people I work with are software engineers. They are all very professional - they all seem happy and the company seems happy with them.
We have an IT department that looks after systems administration - I've only met one of them, once... they are very efficient but hide somewhere and only accept questions via phone or a web-form.
I have friends who call themselves IT consultants - they are so professional that they don't even have to write any software, ever... they make presentations about it and talk a lot. They don't care about being valued because they get paid loads.
Are all three groups IT workers?
Daleks are now so collectable that people are actually stealing them...
Which is why Terry Wogan's commentary is so entertaining, and well worth a watch over dinner and a few drinks.
BTW this years winner was Greece and France were last...