Super artificial intelligence this guy beat Google but maybe not the next time
Andrej KarPathy knows games and artificial intelligence.
His first confrontation was in the year 2011 and artificial intelligence. A group of researchers at Stanford University have just made the world’s best image-recognition software, and he wanted to look at the brain standard image recognition with their digital creations in testing has many advantages.
Software analysis from Stanford University picture library consisting of nearly 50,000 pictures, each picture a kind of fall into ten categories, such as “dog”, “horse” and “truck”. Software the recognition accuracy is approximately 80%. Karpathy conducted the same test, complete with 94% accuracy of artificial intelligence. Karpathy is a graduate student at Stanford University, he thought for a long period of time, humans have this kind of testing is superior to machines. Karpathy said in a reference to artificial intelligence algorithm blog:
“Artificial intelligence algorithm accuracy is difficult to more than 80%, but I guess it was the accuracy can be increased to 85% from 90%. ”
But he’s wrong.
Last year, Google researchers have built a system in more complex ImageNet image recognition accuracy of test at 93.4%. Karpathy, together with some colleagues from Stanford University, and confrontation with this system. But this time, their accuracy at the beginning only at 85%.
2011 test cannot be likened to the ImageNet, but the important thing is: humans in the year 2011 could easily defeat the artificial intelligence software is now dying. Absolutely no chance of winning. Guerlain iPhone 5 Case
This story can be extended very far, from Google to Facebook, from IBM to Baidu, people everywhere brimmed with excitement about artificial intelligence. These firms are invested heavily in depth study, an emerging field, deep learning is carried out according to the accumulated knowledge of the brain pattern modeling. Since 2013, ImageNet race champions are deep learning algorithm, and its voice recognition, video recognition, even in the analysis of financial performance.
University of California, Berkeley, Professor, artificial intelligence expert Stuart Russell says, which shook the field of artificial intelligence, because in the past people have long thought impossible is solved by computer by computer problems all of a sudden.
In other words, the computer also has a lot to learn.
Artificial intelligence training camp
Karpathy and his colleagues finished system by Google is one of the causes of child abuse, ImageNet way of dealing with such things as dog. Karpathy attend the 2011 tests, only one category of dog. But in 2014, ImageNet needs to you 200 breeds of dogs.
In other words, Karpathy must know the Rhodesian Ridgeback and Hungary hunting (this is what dogs … ) The difference between. Guerlain iPhone 5 Case
“When I saw the pictures of all these dogs, I felt, ‘ Oh, no! When in the machine identifies the picture, I was busy trying to confirm the dog’s breed. ”
Therefore, Karpathy participated in the make your own artificial intelligence training camp in this process became the authority identify the small dog breeds. Two weeks later, after more than 50 hours after the training and testing of random click image, he won the machine. Karpathy’s accuracy in testing at 94.9%, 1.7% higher than the Google software.
Human victory again, but winning was not easy. Karpathy says:
“Competition makes me a little exhausted, but I think it is important to know the accuracy of human. ”
Meanwhile, Karpathy and his colleagues hope that the artificial intelligence can be improved. They are working on ways to eliminate defects in the system of artificial intelligence, in an attempt to understand the performance of your computer will achieve human-level at the same time, also trying to analyze their mistakes.
When tested, if there are abstract pictures, Karpathy basically beat the machine. For example, humans can immediately tell if a picture is painted on the bow. Karpathy would recognize the salt shaker on the “salt shaker”, see see things, but computers are not very good at dealing with abstractions.
Computers are not good at identifying 3D images. Computers may be able to identify the Jack Russell Terrier, but to measure its size, or clarify its position relative to the other objects in the same room, is another matter. This is what Google wanted to solve one of the problems, they hope to identify computers like humans in the picture depth and subtlety.
Next, humans can overcome: artificial intelligence?
Bracelet with millet millet’s latest wearable device, with intelligent features such as alarm clock sleep monitoring.
View details of the voting >>
Originally published at givenchycase.wordpress.com on April 17, 2016.