Fulvio Romano
Guest
Amazing. Look what I came across. . .http://thekidshouldseethis.com/post/36119475886
Tequila
Guest
I'm a pussy! ! !
I don't know if you've ever seen these.. .http://www.youtube.com/watch?v=stq3nhxuuem&feature=bfa&list=ple8518e018197fbd9
Fulvio Romano
Guest
I asked a question first. Is training always good?
certainly without training a neural network is useless, but excessive training is likely to lead to "overspecialization". a neural network that works well instead must have good ability to "generalize" the problem. a little like for the human being, he must have a "mindedly open" and be able to understand the phenomenon in its entirety. ok, perhaps "understanding" is a term that should not be used in this context, let's say more than anything that the network must be able to grasp the essence of a problem, even without splitting the hair.
In practice, by submitting the training set to the network for several periods, up to a certain point you get a training, in which the network learns to correctly recognize and process the data of the training set, but also works well on "similar" data. The training is likely to overspecialize the network, i.e. the results of the training set will continue to improve, but the network will begin to understand only those, moving away from the real problem for all points other than the training set.
to avoid this problem you use two sets for training. the training set continues to be used for the updating of weights, but downstream of each training period, the stopping criterion is assessed on the error that the network returns compared to a validation set, different, albeit similar, to the training set. an overspecialized network if on the one hand it continues to decrease the error on the training set, instead it begins to increase that of the validation set, and this is a signal that the network is going off the road.
in the image you should clarify quite well this concept. black balls are those of the training set, while the white ones of the valitation set. the blue curve is returned by a well trained network, which generalizes to the right point and passes close to all points. the red curve instead is overspecialized, it perfectly returns the training set, but completely wrongs the validation set, and then it will also be wrong with all other points of the problem.
certainly without training a neural network is useless, but excessive training is likely to lead to "overspecialization". a neural network that works well instead must have good ability to "generalize" the problem. a little like for the human being, he must have a "mindedly open" and be able to understand the phenomenon in its entirety. ok, perhaps "understanding" is a term that should not be used in this context, let's say more than anything that the network must be able to grasp the essence of a problem, even without splitting the hair.
In practice, by submitting the training set to the network for several periods, up to a certain point you get a training, in which the network learns to correctly recognize and process the data of the training set, but also works well on "similar" data. The training is likely to overspecialize the network, i.e. the results of the training set will continue to improve, but the network will begin to understand only those, moving away from the real problem for all points other than the training set.
to avoid this problem you use two sets for training. the training set continues to be used for the updating of weights, but downstream of each training period, the stopping criterion is assessed on the error that the network returns compared to a validation set, different, albeit similar, to the training set. an overspecialized network if on the one hand it continues to decrease the error on the training set, instead it begins to increase that of the validation set, and this is a signal that the network is going off the road.
in the image you should clarify quite well this concept. black balls are those of the training set, while the white ones of the valitation set. the blue curve is returned by a well trained network, which generalizes to the right point and passes close to all points. the red curve instead is overspecialized, it perfectly returns the training set, but completely wrongs the validation set, and then it will also be wrong with all other points of the problem.
Attachments
panormus
Guest
I chased the girls and built huts on the trees. :biggrin:!!!!
I was eight years old the best I could do was disassemble the remote controlled machines and change the electric scooters to "truck them"
Of course, in the end, I always had some pieces.
PierArg
Guest
the "pericles" of artificial intelligence:
http://www.corriere.it/tecnologia/c...pe_c338ab46-37c9-11e2-94e7-603de4c26bba.shtml
http://www.corriere.it/tecnologia/c...pe_c338ab46-37c9-11e2-94e7-603de4c26bba.shtml
Fulvio Romano
Guest
(I guess you don't need to write the source... do you????? :
I wouldn't worry about machines taking over. At least not after twenty years of the last government. . .
numero1
Guest
ah ah ah ah:biggrin:
As for the phrase you wrote the source:
the villain of matrix
even if I'm sure everyone will understand.
PierArg
Guest
Of course we understood
It was necessary to discover only the den of whiteniglio:biggrin::finger:
MBT
Guest
lightning can only indicate the threshold. You're the one who has to go through it!
MBT
Guest
and fortunately it was not a suppository...
Tequila
Guest
But it's the red one that shows you how deep the whiteniglio's den is..
(read my signature :biggrin
Tequila
Guest
Fulvio Romano
Guest
sfidate questo:http://it.akinator.com/personnages/certainly a neural network, probably a support vector machine (powerful clusterer).
Who knows?
Who knows?
PierArg
Guest
This thing here had made me see my girlfriend who being "ignorant" (in a good way) in this matter believed to be who knows what great invention.
I told her it's nothing but a huge database of characters. As you answer questions, the domain (I had to explain this too:tongue of possible answers shrinks.
Look, the number of questions you ask is never the same.
the more the character is "ambiguous" and the more sw needs elements to make selection.
The only way to screw him is to think of a character not present in the database.
I told her it's nothing but a huge database of characters. As you answer questions, the domain (I had to explain this too:tongue
of possible answers shrinks.Look, the number of questions you ask is never the same.
the more the character is "ambiguous" and the more sw needs elements to make selection.
The only way to screw him is to think of a character not present in the database.
Fulvio Romano
Guest
Of course, but do not underestimate the difficulties of "data mining".This thing here had made me see my girlfriend who being "ignorant" (in a good way) in this matter believed to be who knows what great invention.
I told her it's nothing but a huge database of characters. As you answer questions, the domain (I had to explain this too:tongue
Look, the number of questions you ask is never the same.
the more the character is "ambiguous" and the more sw needs elements to make selection.
The only way to screw him is to think of a character not present in the database.
when you have such a large number of elements (about 10^5 in the case of akinator) and you cannot "tag them" by hand, you do not know a priori if the problem is linearly separable, and if you, with how many size of support, find an algorithm that succeeds to have a rapid descent towards the solution, with few ambiguities and zero errors, it is not at all a trivial thing.
the engine of akinator, "limule", seems to work extremely well, even in the probable case of some misleading elements. who knows how many people will be put to give answers to the caxxo, yet the engine remains robust, means that it corrects the inconsistent inputs.
a polemical comment... If the revenue agency used such algorithms, they would take the evaders without breaking the co...siddetti to me, that I pay all taxes and I must also justify in writing to the tax if towards the bank more than a thousand euros in cash, almost assumed that I am a criminal, guilty until proven otherwise.
PierArg
Guest
without doubt the algorithm that performs the data mining is of considerable complexity and difficulty.
you notice the kind of questions that step by step.
I was referring to people who see akinator as a kind of "indovino"...I had to explain that he already has all the answers.
as if it were an electronic "I guess who", only that has thousands of characters:biggrin:
what amazes is not the fact of guessing characters (after n questions it is also logical that the field shrinks) but the type of questions that asks you (always more specific).
you notice the kind of questions that step by step.
I was referring to people who see akinator as a kind of "indovino"...I had to explain that he already has all the answers.
as if it were an electronic "I guess who", only that has thousands of characters:biggrin:
what amazes is not the fact of guessing characters (after n questions it is also logical that the field shrinks) but the type of questions that asks you (always more specific).
Tequila
Guest
I did a test with a real character but that he would hardly recognize, but instead after a lot of yes and no he guessed it. . .
Exatem
Guest
My daughter plays it every now and then I wanted to try it myself and, that's who guessed me. . .
Fulvio Romano
Guest
The big dog lost a pair of legs...
http://www.youtube.com/watch?v=zkbnfpbv3f0It seems that it will be connor decided to move.
this is the evolution of the big dog:
http://www.youtube.com/watch?v=btrnzlo7m_m
http://www.youtube.com/watch?v=zkbnfpbv3f0It seems that it will be connor decided to move.
this is the evolution of the big dog:
http://www.youtube.com/watch?v=btrnzlo7m_m