Home >

Interview With Professor Pieter Abbeel Of University Of California, Berkeley

2020/10/23 10:37:00 0

ProfessorPieterAbbeelAIButton

How can human beings get along with AI by breaking the limitation of repetitive actions of robots and realizing self-learning, environment recognition and feedback? As a "new type of power", AI and its derivatives will occupy an important place in the future. However, people still have to overcome many difficulties in order to achieve this goal and achieve the goal of wide, universal and rational use of AI.

On October 22, Ernst & young, a third-party consulting agency, issued a report, pointing out that Chinese mainland enterprises hold a relatively cautious and conservative attitude towards the future development of artificial intelligence, and 68% of the enterprises surveyed expect that artificial intelligence will have a greater or even significant impact on the industry in the next five years; under the background of government policy support and industrial upgrading, the artificial intelligence maturity of enterprises is relatively high However, there are still some deficiencies in network security, data management and innovation management.

From 2009 to 2019, the number of investment in artificial intelligence in mainland China continues to rise, with a total investment of RMB 282.7 billion. In the investment of artificial intelligence, computer vision and intelligent robot are the main investment fields.

Pieter abbeel, Professor of UC Berkeley and director of Berkeley Artificial Intelligence Laboratory, has been focusing on robot and machine learning, especially in the field of deep reinforcement learning. He is also the co-founder and chief scientist of covariant.ai (warehouse and factory intelligence and robot automation) and graddescope (artificial intelligence scoring system).

Recently, in an exclusive interview with the 21st century economic reporter, he believed that in the development process from industrial robots to home robots, the biggest change is to change the repetitive robots into robots 2.0, that is, robots who can see and produce feedback according to their situations. Since this year, the robot's function of feedback to the environment has played a practical role.

It is worth noting that when AI really "wakes up", will they dominate the world? How far is AI from the evil that humans fear?

The following is the transcript of the dialogue:

21st century: in the first half of this year, the whole world was affected by the outbreak of new crown pneumonia. What role does AI play in human life? Will it become more and more important?

Pieter abbel: in the context of the virus pandemic, the work that forces employees to gather like logistics infrastructure has become dangerous and important because everyone of us needs to shop online, so we need logistics. Therefore, I hope that large warehouses and other places can be more automated, more secure and reliable. In addition, machine learning also helps to track the development and control of things.

I hope that in the future, AI can help people deal with viruses and provide medical services. At present, we have no obvious progress in vaccine development and drug treatment, but I believe that AI has the potential to make outstanding contributions in medical services, far better than what we have achieved so far.

21st century: the development of AI can be divided into several stages. What stage do you think we are in?

Pieter abbeel: in the early 1950s and 1960s, we focused on search and theoretical reasoning, and found that AI can prove some theories that human beings can't prove yet, and can also become excellent chess players. But it's all based on very clear rules, and you can understand why AI does this.

At the end of last century and the beginning of this century, AI has made great progress in module identification and probability reasoning. At present, thanks to a large amount of data, we have made greater progress in these two areas, such as computer vision. Although this technical problem has not been completely solved, many achievements have been produced in the process and put into practical business activities.

At present, we are in a highly developed stage of module identification. People want to solve a problem, need a lot of data to help computer learning.

For example, if a robot wants to recognize a pedestrian or a vehicle, it needs to learn numerous photos of pedestrians and vehicles before putting it into work. But in the future, robots can learn by themselves, and machines can automatically learn a lot of data, just like newborns, and observe the world autonomously. There is no need for human beings to teach it repeatedly and label a lot of information. Maybe in the next five years, we can take advantage of and see the impact of unidentified data, and then derive more applications.

21st century: ordinary people can not apply deep reinforcement learning in their daily life. What are the most difficult and urgent needs at this stage?

Pieter abbeel: it's true that intensive learning is far from the life of ordinary people. But when we think that robots only need to learn the input image, not the actual action or imitation, technology has made great progress. In the past few years, we have combined deep reinforcement learning with autonomous learning, and found that learning from state and learning from images can achieve the same effect. I think this is an important breakthrough. But it's still not efficient enough and needs a lot of testing.

On the other hand, if I want to implement a system, I don't start with compulsory learning, I start with imitation. I will first collect demonstrations of people driving cars and completing their work, and then let robots imitate them, which is more convenient. After that, we can use compulsory learning to improve and perfect.

21st century: how long is it going to take for us to have a wide range of home robots? In the process of developing home robots, what difficulties and opportunities do we face?

Pieter abbeel: every family has different situations and arrangements. Some families have pets or children, and they often make a mess of the room. Therefore, when I think about designing a robot, I need to consider the structure of the robot's working environment in the future.

We can see all kinds of robots working in the factory, but they just keep repeating the same action.

21st century: AI has been widely used in autonomous driving, medical services and education. What other areas will AI show business potential in the future?

Pieter abbeel: my doctoral advisor said that AI is a new type of electricity, it will be everywhere. I quite agree with him. Based on more data, many people can accurately predict AI, and then play its commercial advantages.

In 2020, the biggest changes I will see are: 1. Language processing. This year, there are predictive language models, such as products developed by Google. These models are much better than before, and people will develop many applications based on this. For example, dialogue robots, sports robots, entertainment robots, they can talk with people fluently, and then improve the ability of machine language transformation.

2、 Robotics. A lot of things can't be done just by repeating actions. You need to observe and feedback. In the next few years, robots will complete most of the human hand work.

In this regard, although autopilot technology is the most commonly discussed technology, I think it is more difficult to develop than people think, because AI can make fatal mistakes, and no company wants to let you know that their system may be fatal, so the reliability of this technology is not enough. In the laboratory, you may develop a very good automatic driving product, but when it is introduced into real life, different vehicles will encounter different drivers, and the system is difficult to achieve high reliability.

What are the challenges facing AI in the 21st century? What can we do about it?

Pieter abbeel: it's true that many people have such questions, because we have seen a lot of cases where AI has not been used properly, such as the failure of detection tools, the police catching the wrong suspects, and so on. Therefore, we should be cautious about the ethical issues of AI. If a person makes a mistake, he just makes a mistake. But if AI makes mistakes, there will be countless mistakes.

The crux of the question is how to measure AI as it makes ethical choices. Although sometimes this measurement is not perfect, at least after the measurement, we can start to study and train it to make better decisions. With the development of technology, there will be good solutions, but there is still a long way to go.

21st century: when AI really "wakes up", do you think they will dominate the world?

Pieter abbeel: I don't have an answer to that right now. In real life, just because a person is smarter doesn't mean he has more power. But if AI is more than 100000 times smarter than humans, maybe AI will dominate the world or make a lot of money. It depends on what we want, whether we want to control them, or when they are smart enough to do whatever we want to do? This is a complex issue.

21st century: are there any measures to prevent this? For example, to set standards for some programs?

Pieter abbeel: there's a human AI co-existence center in Berkeley, and we're also studying how to put AI under control, such as setting a "close button," but maybe a robot will prevent you from closing it through physical or emotional means. So we need an AI system that understands that humans are observing it, but not fully understanding human intentions. We don't want AI to be fully aware of human needs, otherwise it will be dangerous.

You can't give robots very specific, strict goals. Robots need to know that they don't understand everything.

 

  • Related reading

Interview With Zhang Ning, Former Director Of Shanghai Securities Regulatory Bureau

Celebrity interviews
|
2020/10/20 9:56:00
0

Interview With Former Vice Mayor Zhang Hongyi Of Shenzhen

Celebrity interviews
|
2020/10/13 12:23:00
339

Interview With Wu Tianhua, CEO Of Tiger Securities: The Boom Of Chinese Enterprises Going To The United States For Listing Continues, Enabling New Economy Investment Banks To Compete For Incremental Value

Celebrity interviews
|
2020/10/10 10:28:00
129

Ma Yun Is The Last Decade To Promote The Digital Industry

Celebrity interviews
|
2020/10/9 10:44:00
1

ZTE Tu Jiashun: The Intelligent Transformation Of Traditional Industries Takes The "Digital" Train First

Celebrity interviews
|
2020/10/1 14:46:00
2
Read the next article

1000亿美元的半导体收购大战:国际巨头攻守背后的奥秘

高昂的收购计划折射着巨头们的攻守策略,英特尔固守大本营,英伟达全军出击,AMD绝地求生,SK海力士和