The US technology news website BackChannel recently published an article about the artificial intelligence team inside Facebook and its development status.
Joaquin Quinonero Candela is hesitant when he is appointed as the head of the Facebook Application Machine Learning Group (AML) to help deploy the artificial intelligence technology in the world's largest social network. .
Jacqueline Candela, Director of Engineering, Facebook AML Business Unit
Candela is a scientist born in Spain who always calls himself a "machine learner." The reason for hesitation is not because he did not see how much artificial intelligence has helped Facebook. Since joining the social networking giant in 2012, he has seen a shift in the company's advertising business – they use machine learning technology to improve the relevance and marketing effectiveness of sponsored content.
More importantly, he armed his own subordinates with technology in a unique way – even if they did not receive professional artificial intelligence training. Not only that, he also expanded the popularity of machine learning technology throughout the advertising sector.
But he is not sure that the same "magic" can be displayed on a larger scale, because the connections between billions of users on this platform depend on vague values ​​rather than hard data to measure advertising. "I need to make sure that it does have value," he said when he mentioned the appointment.
Despite some doubts, Candela accepted the appointment. Now, although only two years have passed since then, his hesitation at the beginning has become very ridiculous.
How ridiculous is it? Candela gave a speech to a group of engineers at a meeting in New York last month. "I want to make an important statement." He warned, "If there is no artificial intelligence, Facebook can't exist now. You may not realize it, but every time you use Facebook or Instagram or Messenger, your experience. There is a credit for artificial intelligence."
Last November, when I came to Facebook's headquarters in Menlo Park to interview Candela and his team, I was able to see how artificial intelligence suddenly became Facebook's survival nutrients. So far, mentioning Facebook's development in this field, many eyes will focus on the company's world-class Facebook Artificial Intelligence Research Division (hereinafter referred to as "FAIR"), the leader of which is the famous neural network expert Yan · Yann LeCun.
Like competitors such as Google, Weibo, Amazon, and Apple (which is now known for its confidentiality, which allows its scientists to publish research results), FAIR has also become a preferred provider of top-notch AI graduates in short supply. . Computers' advances in visual, auditory, and even conversational abilities have benefited from this brain-like digital neural network, and FAIR is one of the most productive institutions in this area.
But Candela's AML business unit is responsible for integrating FAIR's research results with Facebook's actual products. More importantly, they will also help all of the company's engineers integrate machine learning technology into their work.
Because Facebook is already inseparable from artificial intelligence, all engineers must use this technology.
Put artificial intelligence in everyone's hands
Just two days before I visited Facebook, the United States just ended the presidential election, and the company’s CEO, Mark Zuckerberg, just responded a day ago that those who claimed Facebook’s spread of fake news helped Donald Trump ( Donald Trump) The idea of ​​being elected president of the United States is "too crazy." Because people have been dissatisfied with Facebook's fake news floods, Zuckerberg's comments are nothing to add to the fire.
Although many of the controversies are not within Candela's responsibilities, he knows that Facebook needs machine learning technology to solve the fake news crisis, which is precisely one of his team's responsibilities.
But in order to reassure the public relations staff inside the company, Candela also showed me something else to reflect the work his team is doing. To my surprise, this is actually a bit of a boring trick: it can render a photo or a video in accordance with the unique style of a famous painter. It's easy to remind us of all sorts of gimmicks on Snapchat - turning photos into Picasso-style paintings is no longer a new technique.
“This technique is called neuro-style transfer,†he explains. “It’s a large-scale neural network that can be redrawn into a specific style of painting by training.†He pulled out his cell phone and shot A photo was taken and then manipulated on the screen, and the photo was quickly rendered in the style of Van Gogh's famous painting, The Starry Night.
Even more amazing is that he can also render content into a similar style during video playback. But he said that what really matters is actually invisible to the naked eye: the neural network developed by Facebook can already run independently on the phone.
This is also not a novelty - Apple has previously claimed that it can already do some neural network calculations on the iPhone. But because Facebook doesn't control hardware, they face much more difficulty. Candela said that his team was able to complete this "trick" because they have accumulated a lot of experience - each project can reduce the difficulty of other projects, and each project can make future products more acceptable. Develop similar products with less training – thus speeding up the development of similar projects.
“It took me only 8 weeks from starting the project to public testing, which is crazy,†he said.
He said that there is another secret to completing the task in such a short period of time, that is, cooperation - this is precisely the cornerstone of Facebook culture. Specific to this project, it is precisely because of the easy access to the research results of other business units - especially the mobile department familiar with the iPhone hardware - that they can use the image rendering task that would otherwise need to be completed by the data center. Independent implementation.
From left to right, the engineering director of the AML business unit, Jacqueline Candela, the head of the application computer vision team, Manoha Paruli, the technical product, Rita Aquino, and the engineering manager, Rajan Suba
This technology not only makes it easy for users to shoot "scream" style short films for their friends and relatives, but also makes the whole Facebook more powerful. In the short term, this allows the company to better understand the language and understand the text. In the long run, he can also analyze real-time analysis of what you see and say.
"We are in seconds, even shorter than seconds - and must be done in real time." He said, "We are social networks. If I want to predict people's feedback on a certain piece of content, my system will respond immediately. ?"
Tandrler took another look at the Van Gogh-style self-portrait that he had just taken, completely disdain to cover up his pride. “Being able to run complex neural networks on mobile phones can put artificial intelligence on everyone’s hands,†he said. “It’s not accidental, thanks to the way we display artificial intelligence within the company. ."
“This is a long journey,†he added.
Microsoft veterans show their power
Candela was born in Spain. When he was 3 years old, he moved to Morocco with his family and studied at a French school there. Although he graduated with a high score in the liberal arts, he decided to enroll in a school in Madrid to learn the most difficult subject in his view: communication engineering. This discipline not only needs to master the physical knowledge of antennas and amplifiers, but also has a good understanding of the data. He thinks this is "cool."
Candela is fascinated by the professors who develop adaptive systems. He developed a system that used smart filters to improve cell phone roaming signals, which he now calls "the neural network in the baby phase." He is particularly fascinated with training algorithms and does not like to write a lot of code. A semester spent in Denmark in 2000 further stimulated his interest in this area, where he met the machine learning professor Carl Rasmussen.
Rasmussen studied under the legendary figure and machine learning Geoff Hinton in Toronto. On the eve of graduation, Candela was originally involved in P&G's leadership program, but received an invitation from Rasmussen's Ph.D. program. So he chose machine learning.
In 2007, he worked at Microsoft Research in Cambridge, England. Shortly after joining the company, he learned that Microsoft is holding a competition for all employees: The company is about to launch Bing search, so it needs to improve on keyword search ads - accurately predict when users will click on an ad.
The winning team's plan will be put into physical testing to see if it has the value of the final release. The winning team itself will also receive a free Hawaiian trip as a reward. A total of 19 teams participated in the competition, and Candela's team tied for the first time with another team. He got a chance to travel for free, but because Microsoft has not pushed forward more important rewards, he feels he has been deceived - Microsoft has not tested his plan to determine whether it will eventually be launched as a product.
What happened next showed Candela's resolute attitude. He launched a "crazy movement" and persuaded Microsoft to give him a chance. He has conducted more than 50 conversations within Microsoft and has developed a simulator to demonstrate the superiority of his algorithm. He even found the vice president who was directly responsible for the decision: he took the initiative to sit next to the vice president while eating the buffet, and even grabbed the opportunity to go to the toilet with him to promote his plan. He also broke into the executive's office without prior notice, claiming that his speech must count, and his algorithm is indeed better.
In the end, Candela's algorithm was launched in conjunction with Bing in 2009.
Facebook 20 Building Interior
On a Friday in early 2012, Candela visited a friend at Facebook's Menlo Park Park. To his shock, he heard that the company's employees can test their own projects without obtaining approval from their supervisors. They did just that. So he went to Facebook for an interview on Monday and got an acceptance notice on the weekend.
After joining the Facebook advertising team, Candela's job is to lead a team to show more relevant ads. At the time, machine learning techniques were used. "But the model we used at the time was not advanced, it was too simple." Candela said.
There is also an engineer who joined Facebook with Candela, who is called Hussein Mehanna, who is equally surprised by the company's backwardness in artificial intelligence integration. “When I look at the quality of their products as an outsider, I think everything has taken shape, but obviously not.†Mehana said, “In a few weeks, I told Jacqueline that what Facebook really lacks is a A world-class machine learning platform that works well. Although we have machines, we don't have the right software to help the machine learn as much as possible about the data." (Mehana, who is currently the head of Facebook's core machine learning, is also a Microsoft veteran. - Several other engineers interviewed by this article have the same identity. Is this just a coincidence?)
Mehana’s “machine learning platform†refers to the deployment of a state-of-the-art artificial intelligence paradigm: by virtue of several models based on human brain behavior patterns, this paradigm takes this technology from the last century’s “winter†"At the time, the idea of ​​the early "thinking machine" could not lift people's interest) brought to the recent boom period.
Specific to the advertising business, Facebook needs to let its system accomplish tasks that humans can't reach: predicting in real time and accurately how many people will click on an ad. Candela and his team hope to develop a new system based on the machine learning process. And because the team wanted to build the system in a platform that was accessible to all engineers in the department, they worked hard to ensure that modeling and training were widely promoted and replicated during the development process.
One of the important factors in building a machine learning system is getting massive amounts of data—the more data, the better the results. Fortunately, this is one of Facebook's biggest assets: If more than a billion people interact with your product every day, you can collect a lot of training materials and get countless examples of user behavior.
This also allowed the entire advertising team to develop a new model from a few weeks, turning into several new models every week. And since this will be a platform for others to develop their own products internally, Candela must involve multiple teams in the development process. They divided the process precisely into three steps: "focus on performance first, then focus on usability, and then build a community." He said.
Candela's advertising team has proven that machine learning has brought a huge shift to Facebook. “We achieved incredible success in predicting clicks, likes, conversions, and more,†he said. The next step is naturally to extend this approach to more services. In fact, Le Kang, the head of FAIR, has always advocated the establishment of a department to cooperate with the application of artificial intelligence technology to actual products.
"I very much hope to set up such a department, because you need to organize a group of top engineers, although they do not have to pay attention to the products directly, but they need to pay attention to the basic technology, so that many product departments can use them." Lekun said.
In October 2015, Candela became the head of the newly formed AML team (but only for a while, because he was very cautious, and at the same time retained the position of the advertising department, both need to be both.) He kept with FAIR Closely related, the latter has offices in New York, Paris, and Menlo Park. In fact, no matter where FAIR researchers and AML engineers sit next to each other, it is equivalent to setting up a FAIR office there.
The way the two sides work together can be fully reflected in a product under development: this product can provide a voice description for the photos posted by users on Facebook. In the past few years, training a system to identify objects in a scene and to draw general conclusions has become a standard artificial intelligence practice model. For example, this technique can be used to determine whether a photo was taken indoors or outdoors.
But FAIR scientists have recently discovered ways to train neural networks that can describe almost all the interesting problems in a picture, and determine the subject of the picture by the position of the object in the picture and its relationship to other objects. So to accurately analyze the theme of a photo is a human hug, or someone is riding.
“We show this to AML people,†Lekun said. “They thought for a while, 'You know, this is very useful in one situation.'†They later developed a prototype function. When blind and visually impaired people put their fingers on a photo, they can use the phone to paint the content on the photo.
“We have been communicating,†Candela said when referring to FAIR. “The overall goal is to turn basic science into a specific project. This requires a binder, right? We are the binder.â€
Apply basic research to practice
Candela divides artificial intelligence applications into four areas: vision, language, voice, and shooting. He said that these four areas can contribute to a "content understanding engine." Facebook wants to know how to truly understand the meaning of a piece of content, to judge the subtle intent behind the comment; to understand the precise meaning behind the language; to identify the face of a friend in a fast-moving video; to interpret your facial expression and copy it To the virtual reality.
“We want to implement a universal application of artificial intelligence technology,†Candela said. “The content we need to understand and analyze is exploding, but our ability to add tags and distinguish things has not improved simultaneously.†To solve this problem, Develop a common system that allows the results of a project to be accumulated and assists other teams involved in the project.
Candela said: "If I can develop many algorithms and transfer the knowledge of one task to another, isn't that great?"
This conversion can have a major impact on the speed at which Facebook launches its products. Take Instagram as an example. Since its launch, the photo service has shown users' photos in reverse chronological order. But in early 2016, the company decided to use a correlation algorithm to display the image.
The good news is that since AML has applied this algorithm in products such as News Feed, “you don’t have to start from scratch,†Candela said. “They have one or two engineers who are proficient in machine learning and dozens of teams that are deploying various ranking applications. Contact. You can then copy these patterns and, if you have a problem, communicate with the person in charge of these patterns. It is for this reason that Instagram was able to complete such a major transformation in just a few months.
The AML team has been exploring various use cases, combining their own neural networks with the results of different teams to develop a unique feature for "Facebook size." "We are building our core competencies using machine learning technology while pleasing our users," said Tommer Leyvand, chief engineer of the AML perception team. (He is also from Microsoft.)
A recent feature called "Social RecommendaTIons" is a typical example. About a year ago, an AML engineer talked to a Facebook sharing team product manager about the deep interactions people had when they asked their friends for local restaurant and service advice.
“The question is how to present relevant information to users,†said Rita Aquino, product manager for the AML Natural Languages ​​team. The sharing team has tried to match text to specific phrases. "When you accept 1 billion posts a day, this may not be accurate and may not be widely available," Aquino said.
Rita Aquino, Facebook Technical Product Manager
By training the neural network and then testing the various models with real-time behavior, the team can perceive subtle linguistic differences to accurately determine when users are asking for a meal or shopping advice for an area. This triggers a request to appear in the News Feed stream for the contact. Next, machine learning is still used to determine when others provide useful suggestions and to display the location of the business or restaurant on the map in the user's News Feed stream.
Aquino said that during her one-and-a-half-year period in Facebook, the elements of artificial intelligence that were rarely seen in the product became the technology that was integrated from the initial stage. “People want to interact with products that are smarter,†she said. “Other teams will see the social recommendation feature and our code and ask: 'How can we do it?' You don’t have to be a machine learning expert. Try it based on the experience of your department."
Specific to the field of natural language processing, the team also developed a Deep Text system that can be easily used by other teams. It helps with the machine learning technology used by the Facebook translation feature, which is applied to more than 4 billion posts every day.
In the field of graphics and video, the AML team developed a machine learning vision platform called Lumos. The platform originated from Manohar Paluri, who was a FAIR intern at the time and was responsible for developing a magnificent machine learning project that he called "Facebook's visual cortex" - its purpose is to deal with Understand all the images and video content posted on Facebook.
Manoha Paruli, head of the application computer vision team
In a hackathon event in 2014, Paruli and colleague Nikhil Johri developed a prototype product in a day and a half and showed the results to the enthusiastic Zuckerberg and Facebook. COO Sheryl Sandberg.
When Candela formed the AML business unit, Paruli led the computer vision team with him and developed Lumos to help all Facebook engineers (including Instagram, Messenger, WhatsApp and Oculus) take full advantage of this visual cortex.
With Lumos, “anyone in the company can use the functions of these diverse neural networks and then build various models for their specific scenarios to understand the actual operational results.†Paruli said he also works for AML. And the two teams of FAIR, "The last one can let a person correct the system, retrain it, and then push it forward, without the AML team involved."
Paruli gave me a simple demonstration of the effect. He started Lumos on his laptop and then ran a sample mission: refining the neural network's ability to recognize helicopters. There is a page with a lot of pictures - if we keep scrolling, there will be about 5,000 pictures - there are a lot of helicopter photos, and some things like helicopters. (One is a toy helicopter, and some are objects that float in the air like a helicopter.)
During the training process, Facebook used publicly released images (not including content that is restricted to friends or some users). Even if I am not an engineer, I can't talk much about artificial intelligence, but I can easily find negative examples to train the system to build a "helicopter image classifier."
Ultimately, this categorization step called “supervised learning†may be more automated, as the company is pursuing the holy grail of machine learning – “unsupervised learning†– in this model, the neural network can judge for itself What exactly is in these pictures. Paruli said that the company has made some progress. "Our goal is to reduce human annotations by a factor of 100 in the next year," he said.
In the long run, Facebook believes that the visual cortex will work with the natural language platform to become Candela's so-called universal content understanding engine. "We will eventually undoubtedly integrate them." Paruli said, "At that time, we will directly develop the 'cortex'."
Facebook hopes that the core principles they use in technological advancement can be disseminated outside the company by publishing papers, etc., using this democratization model to spread machine learning technology more widely. “You don’t have to spend a lot of time developing smart apps, and the speed can be greatly faster,†Mehana said. “Imagine the impact of this technology on pharmaceuticals, safety and transportation. I think the speed of developing applications in these areas can be accelerated. Hundreds of orders of magnitude."
Faced with no solution
Although AML has been deeply integrated into the R&D process, giving the company's products visual, analytical and even language capabilities, the company's CEO Zuckerberg also believes that in his efforts to use Facebook to create benefits for society, this Technology will play a crucial role.
In the 5,700-word declaration issued by Zuckerberg, the CEO mentioned “artificial intelligence†or “AI†seven times, all in the context of describing how to use machine learning and other technologies to improve social security and information. Mentioned.
It is not easy to achieve these goals, which is the same reason why Candela was hesitant to initially position AML. If you try to be the primary source of information and build personal relationships for billions of users, even machine learning can't solve all the human problems facing this process. Because of this, Facebook constantly modifies the News Feed algorithm—when you can't really determine it yourself, how can you give the system the best combination through training?
"I don't think there is any solution to this problem," Candela said. "If you show the news randomly, you will find it a waste of time. If you only show news from friends, you will win the whole thing. In the end, you will continue to discuss two extreme situations. Which state is the best between us. We are trying to explore some."
Facebook will continue to use artificial intelligence to solve this problem, which has become the driving force for its development in each field. “There are a lot of researches in the field of machine learning and artificial intelligence that hope to optimize the level of exploration,†Candela said hopefully.
When Facebook is used as the culprit of fake news, they naturally ask the artificial intelligence team to clean all news cancer from the platform as soon as possible. This is a rare full-time action, and even the FAIR team, which has always had a long-term outlook, is involved. Lekun said that the team acts as a consultant.
As a result, with the efforts of FAIR, they have developed a tool to help solve this problem: a tool called WorldVec (vec is the abbreviation of "vector"). WorldVec adds some memory to the neural network, helping Facebook tag all content, such as its source, and who has shared it.
With this information, Facebook can learn about the sharing characteristics of fake news and use the company's machine learning algorithms to eradicate cancer. “The results show that finding fake news is not as difficult as judging what people like best,†Lekun said.
The system previously developed by Candela's team has accelerated the speed with which Facebook has launched these audit products. The specific performance of these products remains to be seen. Candela said it is too early to show how much fake news the company has reduced by using algorithms.
But whether or not these new measures work, these puzzles themselves raise a question: whether the pattern of solving problems with algorithms—further enhanced in the era of machine learning—will inevitably lead to harmful consequences. Obviously, some people think that this has happened in 2016.
Candela denied this view. "I think we make the world a better place." Candela also took the initiative to tell a story. On the day of his interview, Candela called him on a contact on Facebook - the man who was his friend's father, who had only seen one side before.
He saw that the man had sent a lot of content to support Trump and was confused about the content. Kandra later realized that his job was to make decisions based on data, but he ignored important information. So, he sent a message to the man, hoping to talk to him. The contact agreed, so he called.
"This didn't change the reality I was in, but it made me look at things in a very different way." Candela said, "In a world without Facebook, I will never have such a contact."
In other words, although artificial intelligence has become a key element of Facebook and even the survival of this platform, it is not the only answer. “The challenge now is that artificial intelligence is still in its infancy,†Candela said. “We are just getting started.â€
Withstand high voltage up to 750V (IEC/EN standard)
UL 94V-2 or UL 94V-0 flame retardant housing
Anti-falling screws
Optional wire protection
1~12 poles, dividable as requested
Maximum wiring capacity of 16 mm2
T16 Series Terminal Blocks,Barrier Terminal Connector,Polypropylene Terminal Block,Polyamide66 Terminal Blocks
Jiangmen Krealux Electrical Appliances Co.,Ltd. , https://www.krealux-online.com