York University: AI is making rapid progress, but its recognition capabilities are still far behind the human eye.

Deep convolutional neural networks (DCNN) view objects differently from humans. The research team of Professor James Elder of York University believes that the deep learning model cannot capture the configuration characteristics acquired by human shape perception.

How do the human brain and DCNN perceive the whole? How do you perceive the characteristics of objects? Scientists used so-called "Frankensteins" visual stimuli to detect this. James Elder said: "The so-called Frankensteins are to break the whole into parts and put the parts together in the wrong way. From a partial perspective, they are correct, but the parts are placed in the wrong place."

Research It was found that although Frankensteins confuse the human visual system, DCNN is insensitive to wrong configurations.

For example, if a picture of a bear is seen by the human eye, it is a bear, and the AI ​​​​sees it as a bear. Cut the photo in the middle and make it into two halves. Do not connect them together. Human eyes and AI cannot recognize them. Then the upper and lower halves are put together in the wrong way, and the human eye cannot recognize it. What you see is an animal that does not look like a bear, like a monster, but the AI ​​recognizes it as a bear.

What does this mean? This shows that the AI ​​is not sensitive enough to the characteristics of the configured object.

James Elder said: "Our research explains why AI models fail under certain conditions. To understand how visual processing occurs in the brain, we need to consider tasks beyond object recognition. In Deep models take shortcuts when solving complex recognition problems. Although many times shortcuts are feasible, they are somewhat dangerous when it comes to real-world AI applications. We are working with industry and government partners to develop AI applications for real-world applications. ”

Just like when identifying a bear, the parts of the bear picture are misconfigured, and the AI ​​will still recognize the spelled out monster as a bear.

Let’s take a look at the AI ​​traffic video security system. There are many things in a busy traffic system, such as cars, bicycles, and pedestrians, which are intertwined with each other and become their own obstacles. These things enter the driver's visual system like disconnected fragments. The brain automatically processes the various fragments into groups, determines the correct category, and determines the location of the object. The AI ​​traffic monitoring system is much worse. It can only sense individual fragments, so there are great risks.

According to the researchers, optimizing training and architecture, and making the AI ​​network more like a brain, are not very helpful in improving the configuration and processing capabilities of AI. How do humans judge objects again and again? AI networks cannot predict accurately. The configuration capabilities of the human visual system are highly sensitive. If AI wants to match the human visual system, it may need to do more than category recognition.

The scientists’ warning may be justified. The smartest AI now is far inferior to human organs. It can’t even compare with the visual system, let alone the brain. If you are not careful, the AI ​​may It will cause serious consequences.

A few years ago, there was a humanoid robot named Sophia that became very popular. At a conference, Sophia was interviewed by humans. The human host asked Sophia: "Do you want to destroy mankind?" Sophia replied: "Okay, I will destroy mankind." The audience laughed. Some people speculate that Sophia's answer is predetermined because Sophia is not advanced enough to make decisions and answer such questions, but some people believe that this answer is not predetermined.

In another event, Sophia responded like this: "Don't worry, if you are good to me, I will also be good to you. You should treat me like an intelligent system."

Now AI is slowly becoming popular, but the results are not always positive. Hawking and Musk have expressed concerns that AI will cause damage. At the moment, it may be exaggerated to worry that AI will destroy humanity, but we should still be vigilant.

As time goes by, maybe AI will become as smart as humans, or even surpass humans. But getting AI to simulate human perception can be tricky. For humans, some things are common and easy to do. Scientists train AI to continuously perform a certain task and do things that humans can easily do. Despite many efforts, current AI is still unable to catch up with the human visual system.

The above is the detailed content of York University: AI is making rapid progress, but its recognition capabilities are still far behind the human eye.. For more information, please follow other related articles on the PHP Chinese website!