Figure Source: Oriental IC
At the just closed 2022 World Artificial Intelligence Conference (WAIC), many viewers have encountered a magical " stealth jacket ": a T -shirt walked through the camera in front of him, It will "turn a blind eye" to you, on the screen used to demonstrate, only among the pedestrians passing by together, only you are not marked by the green box.
"This means that in the final output report, all your information is not among them." The staff of the Ruilai smart Realai at the scene told reporters that in some special scenarios, people wearing this T -shirt appeared in the camera. Human face will not be specially marked and captured, thus avoiding the comparison.
This is a warning. Ten years ago, at the ImageNet Challenge, the Geoffrey Hinton team used neural network deep learning technology to increase the error rate of picture recognition from about 30%to 16.42%, thereby setting off the wave of artificial intelligence in this round.
, after 10 years of development, the academic community generally believes that from the perspective of application, its technical potential is close to the "ceiling" from the deep learning of data. On the one hand, the essence of deep learning is to use data that has not been processed, and use the "black box" processing method of probability to find laws. This method is essentially not explained, unable to migrate, and requires a large number of identified data. On the other hand, the risk point has also appeared. In addition to the "stealth jacket" that saw at 2022waihc, the face "deceived" the bank certification system for fake people's faces, and has appeared in many actual cases in China. In the fifth year of
World Artificial Intelligence Conference, people began to discuss new topics: When the technical dividend of this round of deep neural network learning gradually topped, how to promote the popularization of artificial intelligence to benefit more industries? Where should the new development of artificial intelligence go?
" big data, big models are of course important, big knowledge is equally important. " Chinese Academy of Engineering and Zhejiang University Professor Pan Yunhe pointed out on 2022WAIC that data and knowledge bipoles will be the fourth innovation of artificial intelligence. direction.
"
Makes a digital person" knowledge "
from birth to now, artificial intelligence has gone through 66 years, Pan Yunhe divided it into three stages. 6 The rules and logic -driven artificial intelligence ; the second stage is the 1960s and 1960s, evolved from logic to artificial intelligence driven by knowledge and reasoning. At that time Extensive human experience and way of thinking are more "people" than they are now, but the knowledge expression at that time was characterized. How to become knowledge of visual and sound signals did not solve. The development of artificial intelligence has entered the third stage. Everything that has happened since then, everyone knows that deep neural network has greatly broke through in visual recognition, hearing recognition, text recognition, and multimedia artificial intelligence, but at the same time, many shortcomings have produced many shortcomings For example, the "black box" that cannot be explained, a large number of data that needs to be marked. "Pan Yunhe said that people often say that AI's logic ability is difficult to train, but in fact, the early AI logic ability is very strong, but the deep neural network technology in this round does not have this ability, so it is necessary to combine the two. It is artificial intelligence that is "commonly driven by knowledge and data". Among them, the opening of the road pioneer is likely to be the multiple knowledge expression of other knowledge such as vision and text, that is, the multi -mode artificial intelligence that has been breaking through in the past two years. Typical cross -media artificial intelligence.
Pan Yunhe uses digital human as the interpretation. It not only shows the person's appearance, movement, perception, and human cognitive ability, but also the personalized data of people. It is the difficulty of the Yuan universe ".
Tsinghua University 's research on "multi -mode learning" has also been carried out for a long time. Yao Qizhi, Dean of the Shanghai Institute of Zhi Zhi Research Institute, introduced that the Tsinghua University Cross Information Research Institute Zhao Xing Research Group is pushing multi -modal learning from theoretical to actual application. At present, AI can be automatically generated according to the dubbing script, which will automatically generate and synchronize the rhythm of the picture. High -quality dubbing. It is understood that this research uses the mouth movement in the video to control the rhythm of generating voice to achieve voice and video synchronization.
is gratifying that in recent years, 's coordinated speed is getting faster and faster. According to the "IT Times" reporter, Tencent has applied multi -modal integration to computer vision research, providing the Bank of Communications with visual AI solutions in multiple scenarios, which can quickly handle the unclear upload of the user, the user's documents are unclear, the user's documents Photo recognition, blurring in data seal, and user documents PS for pseudo -pseudo, thereby improving the efficiency of bank transaction processes and improving users' business experience.
"
autonomous driving cannot recognize the" ice cream barrel "
forward to the new technical direction. It is the goal of 10 years after the development of artificial intelligence. Today, data, computing power and algorithm are recognized as" AI three -piece set ". Among them, big data is the basis of the foundation. After many times, it will be considered correctly by artificial intelligence, but why does it come to this conclusion, the middle process is a "black box". That is not enough data, and more data is needed "feed". As a result, most of them are men. This is because engineers in real life are mostly male. To solve all problems and even new risks, the "stealth jacket" seen by the reporter is a "confrontation sample attack". By adding disturbances to the input data, the system makes an error judgment. Lai Smart Realai partner and senior vice president Zhu Meng told the "IT Times" reporter that the test shows that in the autonomous driving scene, the shape of the car barrel can make the car perception module fail and hit it straight up. Special -made glasses, they brushed dozens of people's face passwords for dozens of commercial mobile phones in minutes. Professional chess players can spend 20,000 years to finish, which means that in some small samples, the identification accuracy will be greatly reduced. In a paper, Tesla failed to identify the case of "people holding a parking logo". Because this scene far exceeded the training database, the system did not know what to do. The topic
"For the future development of artificial intelligence, we must do a lot of work in basic research." Yao Qizhi pointed out that Chinese researchers must "have no self" ecosystems.
Tsinghua University Cross Information Research Institute Gaoyang Research Group achieved breakthroughs in high -efficiency reinforcement learning last year.
ATRAI games are one of the most commonly used performance test standards in the field of learning.In 2015, the algorithm DQN proposed by the Deep Mind team, through 200M frame training data, reached the average human level in ATARI games. However, the Efficientzero proposed by the Gaoyang team only uses 1/500 of DQN demand data, and the same effect is achieved in 2 hours.
Yao Qizhi believes that building an artificial intelligence innovation highland is to obtain the right to speak on the height topic. In addition, in terms of key technologies, even if China is relatively insufficient at this stage, we must try their best to catch up and strive to enter the forefront of the world as soon as possible. In the direction of emerging theories and technical directions, everyone is on the same running line, and they should strive for opportunities and running together worldwide.
Yao Qizhi focuses on two cross -research directions: quantum intelligence and AI+X. On the one hand, prepare as early as possible, as the quantum computer gradually matures, and the algorithm breaks through. Source of original work. For example, artificial intelligence and materials can be studied together, and new materials are created with new materials.
Author / IT Times Reporter Hao Junhui
Edit / Kick Girl
Capture / Ji Jiaying
Picture / WAIC IT Times ml10
