Keep in mind that their result is on a controlled dataset ("labeled faces in the wild," http://vis-www.cs.umass.edu/lf...) for which a lot of training data is available and on which previously proposed systems already perform well.
So this 97% number is a bit of an adventurous extrapolation. Think of it as only polling in NYC and stating that you can predict the result of the next presidential election. The paper was clear on that point, only the summary made it look catchy as usual.
I read the paper and while the approach of learning a representation for faces, and then classifying in that new space whether the face is the same as model is sound, the representation is trained on a closed dataset (the 4m faces from facebook).
So it means that there is no way for the scientific community to check whether the results are correct or not. The results in the paper lack a comparison to a reproducible result, like using the youtube or faces in the wild datasets to train the representation, and then report results given that representation. This way researchers could validate the approach.
I would never have accepted such paper if I were to review it.
Nothing is performed on the fly. It's just another feature extraction and selection pipeline. 1) Deep Neural Networks also save the feature engineering step (for instance http://media.nips.cc/nipsbooks/nipspapers/paper_files/nips26/1210.pdf) 2) If as suggested by the title you are interested by on-the-fly object recognition, look at Tracking-Learning-Detection (TLD) (http://info.ee.surrey.ac.uk/Personal/Z.Kalal/tld.html)
If your results depend on hardware, software and so on, what you are doing is sampling from the solution space. You can then model that distribution and perform significance testing vs that distribution. What is the probability of your result being correct? your result belonging to the true distribution?
Statistics over mathematical proofs. That's what you want to do.
Here is the current structure proposed by the organizers for storing the social network.
struct user_struct {
int user_ID;
char * name;
char * account_handle;
int number_of_BFFs;
user * BFF_list;
int scratch; };
The BFF_list field is supposed to contain the list of friends of a user. The proposed type, user*, suggests that it should be implemented as an array of user. This means that if a user is in your list of friends (stored by value in the array BFF_list), you cannot be in his list of friends unless you both have the same friends. It can only represent non-symetric friendship where each user is involved once in a BFF_list.
What you learn at university is not about technology, it's rather - to be curious and to explore avenues that you don't know: this will help you draw those lines between opposing domains that no body had seen before - to multitask, meet deadlines, and work under pressure (why would you need that?) - to communicate with people who don't know what you are talking about (customers, boss?) - to teach yourself new stuff.
A PhD is about science, not money. With a little bit of luck, you'll be in for a startup project, but otherwise, there is no big money making in science.
On the other hand, it's going to be very interesting. When you write your thesis, you will learn something about yourself: how you can cope with frustration, and how you can do stuff that you though you would never be able to do.
But it's all like professional sports, only a few make it to the top.
Using the very same method, we could leverage moderation logs to predict which comments are informative, funny, deceptive... Please admins, where's the API?
The Samsung Galaxy i7500 is supported by a fork of the fork called GAOSP, available in alpha at http://code.google.com/p/gaosp/. This kind of saves this 1.6-only device.
They gathered year predictions on milestones like "a majority of mobile phones can translate conversations" from 127 researchers from the speech community and compared them to those of the same studiy performed 6 years ago and 12 years ago. The funny part is that the averages slide with time, as if the future was near, but unreachable. Also, "never" was a possible answers, and it often showed up with a majority of votes.
From the presentation: * The future appears to be no nearer than it was previously! * The level of scepticism has remained remarkably stable, but pessimism (realism?) seems to have increased
Did you learn anything new in this interview? Seriously, all that has been discussed and rediscussed on Slashdot. Take the new stuff in science, add your personal preference, et voila!
By the way, at the time I looked into iphone development, there was a non disclosure agreement that prevented you from publishing your sources. Is it still active?
How are you supposed to comply with the GPL under NDA? Does that preclude you from using any GPL code?
That also makes the point that you *need* to pay for a mac and a developer license if you want to compile the application for your iphone. You are jailed once again.
When the event of interest is rare, one can report F-score (or F-measure), which is the harmonic mean between recall and precision. Recall is the proportion of cases that you have found in the population in regard of the actual number of cases, precision is the proportion of the cases that you have found in regard to the number of cases that you believe to have spotted. In other words:
Recall = #(right) / #(truth)
Precision = #(right) / #(hypothesis)
F-score = 2 * Precision * Recall / (Precision + Recall)
There are variants to adjust between recall and precision: http://en.wikipedia.org/wiki/F1_score.
Slashdot Mirror
Say welcome to the internet read over AM radio through a synthetic voice. Question: how do I click this link?
Keep in mind that their result is on a controlled dataset ("labeled faces in the wild," http://vis-www.cs.umass.edu/lf...) for which a lot of training data is available and on which previously proposed systems already perform well.
So this 97% number is a bit of an adventurous extrapolation. Think of it as only polling in NYC and stating that you can predict the result of the next presidential election. The paper was clear on that point, only the summary made it look catchy as usual.
I read the paper and while the approach of learning a representation for faces, and then classifying in that new space whether the face is the same as model is sound, the representation is trained on a closed dataset (the 4m faces from facebook).
So it means that there is no way for the scientific community to check whether the results are correct or not. The results in the paper lack a comparison to a reproducible result, like using the youtube or faces in the wild datasets to train the representation, and then report results given that representation. This way researchers could validate the approach.
I would never have accepted such paper if I were to review it.
Why do you still use language to communicate with your pals? You would be so much more efficient with drawings...
Nothing is performed on the fly. It's just another feature extraction and selection pipeline.
1) Deep Neural Networks also save the feature engineering step (for instance http://media.nips.cc/nipsbooks/nipspapers/paper_files/nips26/1210.pdf)
2) If as suggested by the title you are interested by on-the-fly object recognition, look at Tracking-Learning-Detection (TLD) (http://info.ee.surrey.ac.uk/Personal/Z.Kalal/tld.html)
When will a *coin virtual currency make calculations useful for science? I can't help but feel this whole thing is total a waste of energy.
If your results depend on hardware, software and so on, what you are doing is sampling from the solution space. You can then model that distribution and perform significance testing vs that distribution. What is the probability of your result being correct? your result belonging to the true distribution?
Statistics over mathematical proofs. That's what you want to do.
After looking at a lot of alternatives, I settled on OwnCloud News Reader. The android app is still being polished but the browser interface is great.
I think you've just found how to save rss.
Please don't, people will be so much more creative...
The typedef is on user_struct.
Reading only the article is so outdated...
Here is the current structure proposed by the organizers for storing the social network.
struct user_struct {
int user_ID;
char * name;
char * account_handle;
int number_of_BFFs;
user * BFF_list;
int scratch;
};
The BFF_list field is supposed to contain the list of friends of a user. The proposed type, user*, suggests that it should be implemented as an array of user. This means that if a user is in your list of friends (stored by value in the array BFF_list), you cannot be in his list of friends unless you both have the same friends. It can only represent non-symetric friendship where each user is involved once in a BFF_list.
I would suggest using type user** for this field.
What you learn at university is not about technology, it's rather
- to be curious and to explore avenues that you don't know: this will help you draw those lines between opposing domains that no body had seen before
- to multitask, meet deadlines, and work under pressure (why would you need that?)
- to communicate with people who don't know what you are talking about (customers, boss?)
- to teach yourself new stuff.
I really don't like it when, to make more ad revenue by having me click through, story titles and summaries are abbrev...
A PhD is about science, not money. With a little bit of luck, you'll be in for a startup project, but otherwise, there is no big money making in science.
On the other hand, it's going to be very interesting. When you write your thesis, you will learn something about yourself: how you can cope with frustration, and how you can do stuff that you though you would never be able to do.
But it's all like professional sports, only a few make it to the top.
Using the very same method, we could leverage moderation logs to predict which comments are informative, funny, deceptive...
Please admins, where's the API?
How about using a common random prefix followed by a phrase unique to that system? Like:
Xhk645k_networkaccount ...
Xhk645k_elevatedprivileges
Xhk645k_hrsystems
You only have to remember the random prefix, the second part being much easier to remember.
To security experts: would that be secure enough?
This technology is only intended for one thing : Digital Rights Management. Delivering binaries, that's what they want to do...
The Samsung Galaxy i7500 is supported by a fork of the fork called GAOSP, available in alpha at http://code.google.com/p/gaosp/. This kind of saves this 1.6-only device.
...was performed in the speech community, and it yielded somewhat incompatible results.
http://www.asru2009.org/uploadedimages/talk/rkm_talk.pdf
They gathered year predictions on milestones like "a majority of mobile phones can translate conversations" from 127 researchers from the speech community and compared them to those of the same studiy performed 6 years ago and 12 years ago. The funny part is that the averages slide with time, as if the future was near, but unreachable. Also, "never" was a possible answers, and it often showed up with a majority of votes.
From the presentation:
* The future appears to be no nearer than it was previously!
* The level of scepticism has remained remarkably stable, but pessimism (realism?) seems to have increased
Did you learn anything new in this interview? Seriously, all that has been discussed and rediscussed on Slashdot. Take the new stuff in science, add your personal preference, et voila!
Here, you assume that you need a computer to compile an iphone application.
What about compiling applications directly on the phone? Albeit a little jailbreak, you can already run gcc on it.
By the way, at the time I looked into iphone development, there was a non disclosure agreement that prevented you from publishing your sources. Is it still active? How are you supposed to comply with the GPL under NDA? Does that preclude you from using any GPL code?
That also makes the point that you *need* to pay for a mac and a developer license if you want to compile the application for your iphone. You are jailed once again.
When the event of interest is rare, one can report F-score (or F-measure), which is the harmonic mean between recall and precision. Recall is the proportion of cases that you have found in the population in regard of the actual number of cases, precision is the proportion of the cases that you have found in regard to the number of cases that you believe to have spotted. In other words: Recall = #(right) / #(truth) Precision = #(right) / #(hypothesis) F-score = 2 * Precision * Recall / (Precision + Recall) There are variants to adjust between recall and precision: http://en.wikipedia.org/wiki/F1_score.