News classification
Contact us
- Add: No. 9, North Fourth Ring Road, Haidian District, Beijing. It mainly includes face recognition, living detection, ID card recognition, bank card recognition, business card recognition, license plate recognition, OCR recognition, and intelligent recognition technology.
- Tel: 13146317170 廖经理
- Fax:
- Email: 398017534@qq.com
he 5 major technical problems facing artificial intelligence AI in 2018
he 5 major technical problems facing artificial intelligence AI in 2018
In 2017, artificial intelligence made remarkable progress with the help of deep learning. For example, a robot poker player named Libratus beat three human players in a row to become a "Alfa dog" in the poker world. In the ideal, AI is improving many traditional industries, such as autopilot, agriculture, health care, and so on.
Although the speed of AI's expansion is a little skeptical, of course, there is a bubble in AI, and all kinds of media are also hyping AI. But there are still some wise voices: Elon Musk about artificial intelligence is still unable to do so many things. The most troublesome are the following questions:
Understanding human speech
As usual, machines are better at dealing with text and speech than ever before. For example, the Facebook can stop images from the visually impaired, and the Google mailbox can automatically respond to the mail (based on the discontinuation of the mail content). But the AI system still doesn't really know what we mean and what we really think. Portland State University professor Melanie Mitchell (Melanie Mitchell) said: "we can separate the concepts we learned in different ways and apply them in new situations. But AI can't do that. "
Mitchell defines the defect of today's AI as "meaning obstacle" by mathematician Gian Carlo-Rota. Some leading AI research teams are trying to find out how to climb it.
.
The part of this work is to provide the root of a common sense for the machine and the material world that supports our own thoughts. For example, Facebook researchers are trying to teach AI to understand the ideal through watching video. Others are simulating what we can do with the knowledge of the world. Mitchell has tried to use the system of analogy and the concept of the world to explain what is happening in the picture.
Make robots more like people
The robot hardware has been done quite well, you can spend $500 to buy high-definition camera drone palm size, also carrying the box and the machine has been improved on two legs. But this does not mean that it can be used universally because today's robots lack the brain that matches their advanced muscles.
Letting a robot do anything needs to stop specific programming for a specific task, and he can learn to operate from it. But the process is relatively slow. One shortcut is to let the robot exercise in the imitated world and download the hard - earned knowledge into the physical robot. However, this method has been interfered by the ideal gap. Virtual robots are not always effective in imitating the skills learned in the physical world.
But the fluke is that the ideal gap is decreasing. In October, Google's imitation robot society picked up various objects including belt dispensers, toys and combs. From the experimental report, we saw the desired results.
In addition, the automatic driving company deploys virtual vehicles on the virtual streets in the motorized driving competition to reduce the time and money needed to test under the theoretical traffic and road conditions. Chris Urmson, chief executive of Aurora, a self driving venture, said that making virtual tests more suitable for real vehicles is one of the key points of his team.
Guard against hacker attacks on AI
When the Internet is born, it is accompanied by the problem of peace. We should not expect that the autopilot and the home robot would be different. And in fact, it may be worse: because of the complexity of machine learning software, there are many ways of attacking.
This year, research shows that you can hide a secret trigger in machine learning system, which can induce the system to turn into evil way when it sees specific signals. The New York University team designed a functioning street sign identification system unless it sees a yellow Post-It. Put a convenience sticker on one of Broolyn's parking signs, and the system will report it as a speed limit sign. This may cause problems for autopilot.
This threat is considered to be very serious. The most famous machine learning conference in the world held a seminar on Hostage machine fraud earlier this month. The researchers discussed such issues as "how to make people look very normal but the machine opens up to a special handwritten number." The researchers also discussed the possibility of avoiding such attacks and worried that artificial intelligence was used to defraud human beings.
Tim Hwang of the organizational Symposium predicts that using this technology to dominate people is inrereable as machine learning becomes easier to deploy and more powerful. He says that machine learning is no longer the exclusive degree of a doctorate. Tim Hwang pointed out that the false intelligence campaign launched by Russia during the presidential election in 2016 could be the precursor of AI to strengthen the information war. He said: why do not we see the technology of machine learning in these activities? A particularly effective way of Hwang prediction is to use machine learning to generate false videos and audio.
Where is the real future of AI games?
AlphaGo was also fast in 2017, and in May this year, a more powerful version defeated the Chinese go champion. It is the inventor of DeepMind Research Institute, then cut a version: AlphaGo Zero, this version without the study of human chess, chess can also have special skills, based on this it has learned to play chess and chess in japan.
The result of human and AI games is impressive, but it also hints at the limitations of our artificial intelligence software. Chess, Japanese chess
Although the speed of AI's expansion is a little skeptical, of course, there is a bubble in AI, and all kinds of media are also hyping AI. But there are still some wise voices: Elon Musk about artificial intelligence is still unable to do so many things. The most troublesome are the following questions:
Understanding human speech
As usual, machines are better at dealing with text and speech than ever before. For example, the Facebook can stop images from the visually impaired, and the Google mailbox can automatically respond to the mail (based on the discontinuation of the mail content). But the AI system still doesn't really know what we mean and what we really think. Portland State University professor Melanie Mitchell (Melanie Mitchell) said: "we can separate the concepts we learned in different ways and apply them in new situations. But AI can't do that. "
Mitchell defines the defect of today's AI as "meaning obstacle" by mathematician Gian Carlo-Rota. Some leading AI research teams are trying to find out how to climb it.
.
The part of this work is to provide the root of a common sense for the machine and the material world that supports our own thoughts. For example, Facebook researchers are trying to teach AI to understand the ideal through watching video. Others are simulating what we can do with the knowledge of the world. Mitchell has tried to use the system of analogy and the concept of the world to explain what is happening in the picture.
Make robots more like people
The robot hardware has been done quite well, you can spend $500 to buy high-definition camera drone palm size, also carrying the box and the machine has been improved on two legs. But this does not mean that it can be used universally because today's robots lack the brain that matches their advanced muscles.
Letting a robot do anything needs to stop specific programming for a specific task, and he can learn to operate from it. But the process is relatively slow. One shortcut is to let the robot exercise in the imitated world and download the hard - earned knowledge into the physical robot. However, this method has been interfered by the ideal gap. Virtual robots are not always effective in imitating the skills learned in the physical world.
But the fluke is that the ideal gap is decreasing. In October, Google's imitation robot society picked up various objects including belt dispensers, toys and combs. From the experimental report, we saw the desired results.
In addition, the automatic driving company deploys virtual vehicles on the virtual streets in the motorized driving competition to reduce the time and money needed to test under the theoretical traffic and road conditions. Chris Urmson, chief executive of Aurora, a self driving venture, said that making virtual tests more suitable for real vehicles is one of the key points of his team.
Guard against hacker attacks on AI
When the Internet is born, it is accompanied by the problem of peace. We should not expect that the autopilot and the home robot would be different. And in fact, it may be worse: because of the complexity of machine learning software, there are many ways of attacking.
This year, research shows that you can hide a secret trigger in machine learning system, which can induce the system to turn into evil way when it sees specific signals. The New York University team designed a functioning street sign identification system unless it sees a yellow Post-It. Put a convenience sticker on one of Broolyn's parking signs, and the system will report it as a speed limit sign. This may cause problems for autopilot.
This threat is considered to be very serious. The most famous machine learning conference in the world held a seminar on Hostage machine fraud earlier this month. The researchers discussed such issues as "how to make people look very normal but the machine opens up to a special handwritten number." The researchers also discussed the possibility of avoiding such attacks and worried that artificial intelligence was used to defraud human beings.
Tim Hwang of the organizational Symposium predicts that using this technology to dominate people is inrereable as machine learning becomes easier to deploy and more powerful. He says that machine learning is no longer the exclusive degree of a doctorate. Tim Hwang pointed out that the false intelligence campaign launched by Russia during the presidential election in 2016 could be the precursor of AI to strengthen the information war. He said: why do not we see the technology of machine learning in these activities? A particularly effective way of Hwang prediction is to use machine learning to generate false videos and audio.
Where is the real future of AI games?
AlphaGo was also fast in 2017, and in May this year, a more powerful version defeated the Chinese go champion. It is the inventor of DeepMind Research Institute, then cut a version: AlphaGo Zero, this version without the study of human chess, chess can also have special skills, based on this it has learned to play chess and chess in japan.
The result of human and AI games is impressive, but it also hints at the limitations of our artificial intelligence software. Chess, Japanese chess