Shulou(Shulou.com)06/03 Report--

In today's day when everyone may be fooled, why don't we change the object of the game and see if the machine is likely to be fooled?

In the 50 years of research and development of artificial intelligence, we have been trying to make machines understand the way people perceive the world. Whether it is brain-like computing that has not achieved a major breakthrough, or artificial neural network that imitates the mechanism of human perception of the outside world, it is essentially a variety of simulations of human behavior.

Of course, we also know that although these imitations have made breakthroughs in application, they are very different from human cognitive style in essence.

Take vision as an example, human cognition of everything comes from a comprehensive perception. In the big Pok é mon movie that went viral on social media, people were surprised to see the 3D version of Pikachu-Pikachu is hairy?

In fact, this reflects a very interesting point of human beings, based on comprehensive knowledge, the five senses of human beings are connected, so we can mine the association from the limited information and build up cognition of unfamiliar things.

When we simply have furry Pikachu, we immediately think of the furry touch and think that it looks like a big mouse and is not cute.

Machine thinking fooled at will

In contrast, the cognitive style of machine vision is relatively isolated. After the establishment of the classifier, layers of neural networks are organized to deal with the pictures in layers to identify whether the picture is a bridge, a monkey or a big tree. The final conclusion is that this picture has a 97% chance of being a bridge, a 2% chance of being a monkey, and a 1% chance of being a big tree.

For humans, we may think of monkeys as orangutans because of the lack of knowledge at the bottom of our minds to distinguish between orangutans and monkeys. But we will never confuse such irrelevant things as bridges, monkeys and trees.

But for machine vision, it is different. In the "eyes" of the machine, all images are the arrangement and combination of pixels. For us, the difference between monkeys and trees is the difference between mammals and trees. But for the machine, there is only a digital dividing line between the monkey and the tree.

This leads to the fact that machine vision can be fooled by "pertinence" so that the output of image recognition is completely wrong, which is what we often call the fight against generating samples.

For example, if you move the pixels of a picture slightly, there is no difference between the two pictures in human eyes, but under the machine recognition logic, the machine may recognize the monkey as a big tree.

For example, we have introduced the "psychedelic sticker"-the classification features of a certain object are highly condensed into a very small pattern and "pasted" on other pictures. The output of image recognition is based on the ratio of several results. Before the sticker is affixed, image recognition may clearly analyze that the picture has a 98% chance of being a monkey. However, after pasting a sticker with highly condensed features, the result of image recognition can be changed immediately.

It may be easier for humans to understand machine thinking than they think.

In this way, although machines have been trying to imitate the human way of thinking, the end result is that we and machines do not understand each other.

What if we change the angle and let humans learn the way machines think?

Johns Hopkins University recently conducted such an experiment.

To understand why machines can see changes that humans can't see, Johns Hopkins University has launched a series of experiments in which human volunteers find "machine mistakes" in pictures.

For example, give volunteers a noise map or a seemingly meaningless pattern, allowing human volunteers to identify which object the picture is more like, and confuse the results given by AI. Another example is to give human beings a few mottled digital patterns and let human volunteers choose what wrong numbers AI may have mistaken for the pattern.

Shockingly, in most experiments, human volunteers quickly recognized AI's mindset by intuition. In the above picture, for example, 81% of human volunteers have accurately discovered the mistakes that machines can make. Of the 1700 volunteers in 48 experiments, people chose the same answer as a machine 75% of the time, while only 2% never chose the same answer as a machine.

This amazing experimental result tells us that even if we put aside the basic human cognitive logic and apply the first reaction of the subconscious, we can get the same results as the machine. It seems that the depth of refining human experience in the neural network black box may be much deeper than we can imagine.

Antagonistic images: when humans begin to understand machines

The significance of this experiment, of course, is not to tell you that we are all robots, but to teach us how to deal with the mistakes that machine vision may make.

Under the inherent disadvantages of the cognitive model of machine vision, there are likely to be many loopholes for exploitation. For example, when autopilot recognizes a traffic sign on a highway, a small sticker may give the visual system the illusion that the traffic sign can be identified as a sudden pedestrian.

In order to solve this possible situation, it may be a good idea to understand machine thinking and reverse disassemble it.

From the experiment, we can see that even if there are illogical patterns and textures, human beings can recognize and process these pictures at a higher level through the ability of association. In the experiment, humans and machines entered the same situation and made multiple-choice questions in the face of several classification options. And this kind of classification choice may be the key to getting the machine into error.

Another point is that humans can clearly identify adversarial images. For example, deliberately mottled to make it difficult to recognize digital patterns, as well as very distinctive gaudy stickers. Then it shows that these antagonistic patterns also have their own "antagonistic characteristics".

Although in the face of this situation, the best way is to abandon the convolution neural network and apply cognitive computing to truly allow artificial intelligence to be identified on a more complete worldview. But before we make a breakthrough in cognitive computing, we may be able to find some alternative solutions.

For example, in some key situations such as security and self-driving, we can specially generate confrontational samples for human labeling, and add an "antagonistic sample" to the conventional classification of cars and pedestrians. so that the convolution neural network can identify these disruptive patterns.

For the situation in which machine cognition can be changed by moving a few pixels, it is likely to occur errors caused by compression on the way of data circulation, we can look for compression laws that change the results of machine cognition, and reverse the occurrence of errors from the source.

On April Fool's Day, the news that "humans and machines have very similar modes of thinking" sounds like a joke. But perhaps the "human consciousness" that we are proud of is not that complex at all and will eventually meet machine thinking in an unexpected way.

But before we meet, we should still use the similarity between human thinking and robot thinking to solve real-life problems.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.