Shulou(Shulou.com)06/02 Report--

Https://www.toutiao.com/a6689219578108576263/

2019-05-10 10:28:40

For more highlights, please follow the official Silicon Valley Insight website (http://www.svinsight.com).

In 2017, Li Feifei, a professor at Stanford University who launched the ImageNet competition, announced that this year will be the last challenge. Because the error rate of machine recognition of photos has been reduced to almost the same as or even higher than that of humans.

In 2018, the Turing Prize of the Nobel Prize in computer science was announced to be won by the three giants of deep learning: Yoshua Bengio, Yann LeCun and Geoffrey Hinton.

From machine image recognition surpassing humans, to deep learning recognized by the Turing Award, but this is not the end of artificial intelligence.

How do Silicon Valley giants use artificial intelligence? Will there be "super intelligence" in the future AI world? How should small businesses deal with it? Silicon Valley Insight takes you to hear what Daniel Professor Pieter Abbeel of the computer Science Department of the University of California, Berkeley has to say.

From supervised learning to reinforcement learning: targeted AI will have a great impact in the future

Pieter Abbeel holds a doctorate in computer science from Stanford University and studied under Wu Enda. Since 2008, Pieter Abbeel has been a professor in the Department of Electrical Engineering and computer Science at the University of California, Berkeley. He also worked as a scientist at Open AI and co-founder of Gradescope, an education company that changed homework for AI, which was acquired in 2018.

(in Professor Pieter Abbeel's class, the copyright belongs to the Entrepreneurship College)

Professor Pieter Abbeel starts with the three main ways to realize machine learning: supervised learning (Supervised Learning), reinforcement learning (Reinforcement Learning) and unsupervised learning (unsupervised Learning).

The so-called supervised learning is to "feed" a large number of images to the neural network, and the neural network learns from these marked images, and finally finds the object that minimizes the difference with the markers.

In fact, Professor Geoffrey Hinton of the University of Toronto, known as the "father of deep learning", won the 2012 ImageNet competition by proposing a deep convolution neural network structure, AlexNet, 41 per cent higher than the second place at that time. What they use is supervised learning.

Later, many people saw the 2012 competition as a catalyst for today's wave of artificial intelligence. By 2014, it can be said that almost all the high-scoring contestants have used deep neural networks.

Nowadays, artificial intelligence has gone from simply identifying objects in images, to describing images, and then to visual Q & A challenges. It can be said that the machine not only has the ability to understand the image, but also can describe the content of the image in human language, and even about any picture, the machine can ask and answer itself. With each step, the ability of artificial intelligence is becoming more and more powerful.

So, what do you need if supervised learning is done well? Professor Abbeel pointed out that first, a large amount of tagged data, second, a large amount of computing resources, and third, a good AI team. But then comes the challenge: companies need to have a lot of tagged data, which is a "labor-intensive" industry.

This is why Professor Abbeel pointed out that reinforcement learning will be more exciting in the near future.

The main difference between reinforcement learning and supervised learning is that supervised learning is more like having a teacher standing next to you supervising you, but the teacher will know that all the answers are right or wrong. Reinforcement learning is a reward feedback signal to the learning object (that is, the machine). Just like teaching children to walk or puppies, if you do well, you will be rewarded, and if you don't, you will be punished.

One of the most typical examples of reinforcement learning is the AlphaGo that beat two human go masters Ke Jie and Lee se-dol.

Professor Abbeel said that reinforcement learning is currently mainly used in games and robot training to solve common decision-making problems. For example, Berkeley has a robot that simulates human learning and walking, which may fall many times in the learning process, but once it does not fall, the score is even higher, and the result of no fall will be backpropagated to the neural network to enhance the learning results. After about 2000 studies, the robot can learn how to "run".

For Professor Abbeel, he thinks the use of domestic robots in reinforcement learning is even more exciting. His main research direction is to apply deep reinforcement learning to robots.

Since 2010, he and his students have programmed Berkeley's BRETT (Berkeley Robot for the Elimination of Tedious Tasks) robot to pick up towels of different sizes, figure out their shapes and fold them neatly. Now it has been developed to fold clothes, organize toys, cutlery and so on. How many family dreams! But the robot costs 400000 dollars. )

According to Professor Abbeel's prediction, robots will face two new waves of automation in the future:

The first wave: the robot with long eyes. He said that when robots have visual ability, they will adapt to many new scenes and will be able to accomplish more tasks.

The second wave: "teachable" robots, that is, through human demonstration teaching, let robots learn from them. Accomplish tasks that cannot be programmed to teach robots. "what really helps teach robots new skills faster is the use of teleoperation and virtual reality," says covariant.ai, a company that Professor Abbeel now starts with several students.

So when will such robots be rolled out on a large scale to provide automation solutions for manufacturing, warehousing, logistics and other industries? Professor Abbeel expects to land within five years, mainly because the current technology is not yet fully mature, but it has gradually become a reality in the laboratory.

As for unsupervised learning, the main difference between unsupervised learning and supervised learning is that there is no pre-tagged data for training examples, and the machine automatically classifies or groups the input data. Professor Abbeel likens it to teaching a child what a car, a truck or a bike is. You only need to teach it once, without having to tell him again and again. However, there is a lack of unsupervised learning applications at present, and this field will be the place where major companies will gradually make efforts in the future. For example, Yann Lecun, the chief artificial intelligence scientist of Facebook, won the Turing Award for his research on unsupervised learning.

How do Silicon Valley business giants use AI? Give priority to in-depth supervised learning

Is artificial intelligence widely used in the business field at present? Professor Abbeel believes that for business cases, it is mainly deep supervised learning, and it is very effective, creating a lot of possibilities.

The first very important application scenario is: automatic prediction ability. Professor Abbeel even quoted his doctoral supervisor Wu Enda as saying: "I think it will be difficult to cite an example in the next few years that cannot be subverted by AI."

Now, let's take a large group of technology companies in Silicon Valley as an example to see how deep supervised learning is applied in daily life:

For example, food review site Yelp (similar to the American version of "Dianping"), imagine how many photos users upload in a day? In 2016, the number ranged from 100000 to tens of millions.

Just imagine, how to show more beautiful photos to other diners so as to attract more diners? This is where artificial intelligence works.

As for the data mentioned earlier, it is not realistic for Yelp team members to label photos one by one that look good or not. First, it is too subjective; second, it is too wasteful of labor. So Yelp's team thought of the training database. They found that the parameters that determine the quality of a photo include whether to shoot with a digital SLR and the EXIF data that records the attributes of the digital photo and the shooting data.

Therefore, after using deep learning to build a photo scoring model, the quality of photos uploaded to Yelp is now a high point of view. In the following two sets of pictures, does the bottom picture look more appetizing than the one above?

(above: before AI screening, below: photos uploaded after AI screening)

If Yelp is directly oriented to consumers, then business-oriented businesses also apply. Another example of Professor Abbeel is Salesforce, a famous global customer relationship management (CRM) software service provider.

Everyone who comes to San Francisco is probably impressed by the tallest building in the city: yes, it is now the tallest building in the United States, Salesforce's new office building, Salesforce Tower, which is 326m high and has 61 stories, almost as high as the third ITC in Beijing. Every year when Salesforce holds a company celebration, there is a traffic jam in half of San Francisco.

(the picture comes from the Internet, and the copyright belongs to the original author)

The SaaS company, which customizes customer management service systems for customers on demand, applies artificial intelligence to where it is most needed: marketing.

Just imagine, do you receive a lot of subscription emails from merchants every day? But do you open or unsubscribe, buy or browse after you open it? for merchants, mastering such customer dynamics is very important to the final sales success. Because it depends on the merchant to track those customers who are more likely to buy, thus allowing the sales team to prioritize their work.

Today, artificial intelligence can target some customers or retain those users who may be lost by tracking the interaction of each customer in the customer relationship management system.

When it comes to medicine, supervised learning plays an equally important role.

Today, the level of artificial intelligence testing for pneumonia through chest X-rays has surpassed that of professional radiologists. Last year, Wu Enda's team published a deep neural network-based CheXNeXt model that can diagnose 14 diseases, including pneumonia, pleural effusion and lung masses. The diagnosis of 10 of these diseases is comparable to that of human radiologists on AI, and one is higher than that of humans. But it is worth mentioning that the diagnosis speed of AI is 160 times faster than that of humans.

Professor Abbeel said the US Food and Drug Administration (FDA) also approved the first artificial intelligence device last year to detect specific eye problems associated with diabetes. The device scans the retina of the eye to determine whether diabetes has progressed to a certain extent that affects vision, so that patients can be reminded to see an ophthalmologist in time.

(in Professor Pieter Abbeel's class, the copyright belongs to the Entrepreneurship College)

So, will artificial intelligence surpass human intelligence and achieve super artificial intelligence? In Professor Abbeel's view, it is true that the hardware part is getting closer to people, but the software part is still temporarily missing. But one thing is certain: if super-intelligence is achieved one day, it will not want humans to unplug the power. This needs to make sure that the first attempt at "super intelligence" is the right one, otherwise there may be problems that are inconsistent with human values. Even he himself makes it hard to imagine how the real scene will happen.

Competition of small companies in the AI era

Perhaps many people will think that the current era of AI has been dominated by big companies, whether it is AI top talent or data. But in Prof Abbeel's view, this is not the case. The key for any company to win in the face of competition in the AI era is to grasp the following points:

First, data is the key. Although small companies have much less data than large companies, as small businesses in a certain specialized field, they need to store more data no matter what they do. For example, for example, if a customer service makes a phone call, can you record the call? Is it possible to write down the words of online customer service? Whether it is speech recognition or character recognition, this is an area where artificial intelligence can be fully trained.

Second, to understand what role human beings can play in the future. Once an enterprise has data, it should be trained to do what can be done with those people in half a second. Or maybe people combine with AI and do it together. Therefore, no matter what the data is used for, it requires the management of the enterprise to think clearly about the role of people in it.

For small businesses, pay more attention to what they do repeatedly every day. "as long as you can train a neural network to do, you need to pay attention to it."

Entrepreneurs or entrepreneurs, do you have a feeling that AI is not so far away from business, but can be landed? The above content is only a wonderful excerpt from the course "how artificial Intelligence will change Business" taught by Professor Pieter Abbeel to the students of the Silicon Valley visiting group of the founding college EMBA.

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.