Top AI can't recognize these simple images

Note: As more and more things depend on more and more elusive AI, it becomes more and more important to find the defects of the latter, which is more and more necessary in this paper.

What's the pattern? It's a simple yellow-black bar. But if you ask the most advanced AI, the answer will be the school bus, and 99% sure. But the AI is wrong.

It is true that the computer image recognition technology is now very advanced. For example, the following image AI cannot recognize a Chihuahua dog in a Mexican hat (some people may not recognize it), but at least it can be identified as a dog with a wide-brimmed hat. But a recent study by Wyoming University's evolutionary Artificial Intelligence laboratory suggests that these AI may not always be so bright, as the opening example shows, the most advanced AI take these randomly generated simple images as parrots, table tennis rackets, bagels or butterflies.

This discovery compels us to acknowledge an obvious but important fact: that computer vision is not the same as human vision. And since we increasingly rely on neural networks to train computers to recognize images, what the computer thinks is not even known to us.

Using evolutionary algorithms to deceive AI

A random image generated by an evolutionary algorithm, the text underneath the image is an object identified by AI.

One way to figure out the cleverness of these self training algorithms is to look at their stupidity. The researchers decided to see if the most advanced image recognition neural networks were susceptible to active error messages. For example, if these neural networks recognize cats, do they also recognize other things as cats?

To this end, the researchers used evolutionary algorithms to generate random images as visual decoys. They first use the program to generate an image, and then make a slight change to the image. The original image and a slightly modified diagram are then presented to the neural network based on imagenet training. If the modified diagram is considered by AI to be closer to an object than the original, the researcher retains the modified diagram and repeats the process. Otherwise, rewind and try again. The result is that the best people survive-or that the computer's most recognizable image survives (rather than the most appropriate).

Finally, the technique generates dozens of images of the true reliability of the neural network over 99%. They are blue-orange wavy lines, some yellow-black stripes and so on, but in AI eyes, they are starfish (star fish) and school bus (parochial).

Black box Pairing

Why is Ai fooled? Some cases are understandable. For example, if you squint, the school bus (parochial) does look like a stripe between the yellow and the black. Similarly, a randomly generated "monarch butterfly" does look like a butterfly's wing, and the image considered a "ski mask" (ski mask) does look like an exaggerated face.

But the researchers also found that AI is often fooled by pure static images. After using a slightly different evolutionary technique, the researchers produced another set of images (pictured below). These images are almost identical in human eyes, like images on a broken TV. But in the most advanced neural network eyes, these are Centipede (centipede), Cheetah (Cheetah), Peacock (Peacock) and so on.

In the eyes of researchers, neural networks seem to form a variety of visual cues that help identify objects. Some of these clues are familiar to humans (such as school bus examples), and others are not. The example above shows that at least at some point these clues are fine-grained. Perhaps after training, the neural network sees a series of "Green Green, purple, green" pixels as a pattern that the peacock sees. So when a randomly generated image produces exactly the same pixel series, AI treats it as a peacock. This means that AI may be able to infer a number of clues for each object, and that these clues are sufficient to identify an object.

Of course, the human crafting of these images to fool the AI also illustrates the problem that the size and complexity of neural networks are beyond the scope of human understanding-even if we know that AI can recognize images, but they do not know how to recognize them.

The purpose of this research is to deduce the AI model by reverse engineering, and to find out the learning idea of AI. Although still not very much, but the recent two years of black box research has been able to glimpse.

Is there a problem with Ai's vision?

When the researchers presented the findings to the Conference on Neural Information processing systems, the experts formed a stark and divided opinion. A group of people of a slightly older, more experienced field, they think the result is fully understandable. Others are relatively young, and their attitudes toward the results are confusing. At least in the beginning it was surprising that the powerful algorithm had completely mistaken the results. Keep in mind that although these people are slightly younger, they are all people who have published articles at the top AI conventions.

In the opinion of Clune, the head of the research team, this reflects generational shifts in the field. A few years ago, the AI people were developing AI. Now, neural networks are so advanced that researchers can use them.

It's not necessarily bad to take it. But as more and more things are built based on AI, it is increasingly important to discover AI flaws. It's harmless to mistake some random pixel for some kind of animal, but it's very serious if AI makes certain pornographic images slip through the safe search filter. Clune hopes the study could inspire other researchers to take into account the overall structure of the image in the algorithm. In other words, make computer vision more like human vision.

The study also prompted us to consider other forms of AI vulnerabilities. Facial recognition, for example, relies on similar techniques, so there are similar flaws.

If computer vision is focused on local features, perhaps installing a 3D-print nose can make the computer think you're someone else. Wearing a mask will allow you to disappear from the surveillance system. The wider the application of computer vision, the greater the potential for such pitfalls.

But in a broader sense, the warning from this study is that we are entering the era of self-learning systems. Now, we still control what we have developed. But as AI continues to develop itself, one day we find it not surprising that we don't understand AI. "What computers do is no longer written by people writing code," Clune said. "This is almost an interaction between the economies of scale resulting from the intelligent appearance." "We are certainly not wasting time on this intelligent use. But it is not clear whether we fully understand it when we do so.