Knowing the mechanisms of artificial intelligence do you trust its decision more

Source：Updated：2017-07-10 08:34:47

Hey, AI, what do you think you're looking at? Why would machines with learning algorithms be tricked into recognizing things that don't exist? This is becoming increasingly important as products such as driverless cars emerge. Now, we can see the brain of a machine through a test that lets people know which parts of an image ai is looking at. 
 
Artificial intelligence is completely different from humans. Brown university in Rhode Island, said Chris in even the best image recognition algorithm, will also be cheated, such as white noise image recognition into a robin or cheetahs, this is a big problem. He said that if we don't understand why these systems can make stupid mistakes, so when we give life to artificial intelligence, he should think twice before you do, such as prudent considering whether development of driverless cars. 
 
So Grimm and his colleagues created a system to analyse an artificial intelligence in the process of identifying an image, to monitor what part of the image it was looking at. Similarly, for a document classification algorithm, the system shows which words the algorithm USES to determine which category a particular document belongs to. 
 
snooping 
 
This is a very useful way to learn more about artificial intelligence and how it's learned, says du meut erhan, a researcher at Google. Grimsman's tools provide a convenient way for people to check whether an algorithm correctly gives the correct answer, he says. 
 
To create his attention tracking tool, grime installed a second artificial intelligence on the ai he wanted to test. This "outsourcing AI" replaces a white-noised image to see if this affects the judgment of the original software. 
 
If the replacement part of the image changes the result, the area of the image is likely to be an important area of the identification process. The same applies to language. If changing a word in a document makes it different for a document, it indicates that the word is critical to the decision of ai. 
 
Grimm tested his technique on artificial intelligence, which can be classified into 10 categories, including planes, birds, deer and horses. His system maps out the line of sight of artificial intelligence in the classification. As it turns out, ai has learned to break things down into different elements and then search each element in the image to confirm its identification. 
 
Identification horse head 
 
When he saw the horse's image, for example, in the analysis shows that artificial intelligence first has carried on the close attention to the leg, and then search in the image it thought might be the first place - it does not know in advance the horse's head in place. Artificial intelligence takes a similar approach to images that contain deer, but in these cases, it searches for antlers. Artificial intelligence almost completely ignores many parts of an image that do not contain information that helps categorize. 
 
Green and his colleagues also analyzed the ai that can play "Pong" (a video game) after a workout. What they found was that it almost ignored everything on the screen, just keeping a close eye on the two slender moving columns. Artificial intelligence doesn't pay much attention to some areas, so that when the column is removed from its intended location, it mistakenly assumes that it is watching the ball rather than the little column. 
 
Mr Grimm thinks his tools can help people figure out how artificial intelligence can make decisions. For example, it can be used to check the algorithms that detect cancer cells in a lung scan, making sure they don't look at the wrong part of the image, but just find the right answer. "You can see if it doesn't notice the right thing," he says, but first, Grimm wants to use his tools to help ai learn. 
 
If artificial intelligence doesn't pay attention, it allows the ai trainer to direct their software directly to the relevant information.