The future of "machine-depth Learning": Understanding human emotions

The concept of "machine learning" has been a concern of the scientific community since the 50 's. In recent years, "deep learning" has gradually become a new field in machine learning research, whose motive is to establish and simulate the neural network of human brain to analyze and learn, and imitate the mechanism of human brain to recognize the data of image, sound and text.

The latest development of "machine depth learning" technology is summarized by the Internet edition of the American Science and Technology media, Wired magazine. The following is the main content of the article.

In the eyes of Quoc Le, the world is made up of a series of numbers. "A digital photograph is actually a number," he says, "and if you divide what people say into individual phonemes, they can also be compiled into numbers." "If you follow Quoc Le, you can put these numbers into machines, machines can read photos and people say things like Facebook can recognize your face, and Google can understand what you're saying."

But Quoc wants to go further, hoping to develop a technology that translates whole sentences, entire passages, and various types of natural languages into digital or other vectors, so that computer scientists can get the information that people see and hear. At the same time, Quoc le is still exploring ways to make machines understand people's views and emotions.

Although such technology is still in its infancy and there is a long way to go in the future, Quoc le to have more resources for its deployment than its peers. Quoc Le is a member of the "Google Brain" (Google Brain) project, which is mainly engaged in research in the field of "machine depth learning", a form of artificial intelligence that uses machines to simulate the human brain for data processing.

Quoc, 32, has been working on speech recognition at Google, such as the voice recognition feature of the Android system and the automatic tagging of web images, both of which require "deep learning" support.

In addition to Google, internet giants such as Facebook and Microsoft are also using "deep learning" technology, and Baidu recently talked publicly about using the technology to provide customers with more accurate advertising push services. But Quoc wants to push the technology to a wider range, including natural language understanding, robotics, and web search.

Quoc Le recently developed a "deep learning" technology, able to identify how the different words on the network are related, Google in its own "knowledge map" into the technology, so as to help its knowledge of search results systematization, so that each keyword can obtain a complete knowledge system.

Once troubled

Quoc was first exposed to artificial intelligence in the 90 's, but it really bothered him, because the machine learning system relied heavily on the manual input of Engineers, and although the machine had a certain degree of understanding, it needed more cumbersome operations to do so. For example, the machine was unable to identify the photos without tagging them.

"We've studied a lot of data without labels," Quoc Le said he worked with the "Google Brain" project's founder, Andrew Ng, at Stanford University to study AI, "If in the future we can find a feasible algorithm for the machine to identify the data without the label, That would have the potential to change the entire computing industry, after all, and now most of the data on the web, such as Facebook, Twitter and Google, are not labeled. ”

This is precisely the future of "deep learning" technology to achieve the goal. Using tens of thousands of of computers to simulate neural networks in the human brain through software, it allows machines to acquire similar learning capabilities, such as in some cases the ability to automate learning without tagging data.

Google's cat face recognition is actually a typical case of "deep learning" technology, but after three years of research and development, the project still hasn't made much headway. At the same time, most business depth learning systems are still more dependent on manual monitoring. "Although the usefulness of face recognition is low," Wunda said, "but I think the technical representative is a direction for further study in the future." ”

The challenge of language

Another challenge that "deep learning" technology needs to face is the recognition of natural languages. Human language contains a great deal of subtle information, and so far the scientific community has not yet found a way to identify these subtle messages. For example, a similar vocabulary, in different contexts or mood will have different meanings, most AI systems are currently unable to distinguish this information. "Machines are very good at handling data, but they can't cope with language symbols," Quoc le said, "because language is a highly symbolic thing." ”

The key to identifying a language is to find a way to translate symbols into numbers. "We have not yet found a way to translate language concepts into mathematical structures that machines can handle," Quoc le said, "but with the help of Word2vec tools, we have made some progress in this area." Hopefully in the future our machines will automatically identify the vast amount of information posted on the web. ”

"People will not be able to monitor machine learning anytime, anywhere," said Richard Socher, a Ph. D. at Stanford University with Quoc le Richard Soche, "We hope to combine supervised and unsupervised learning in the future, So that the machine can do many things that are now unimaginable. ”

Quoc Le recently published an article on the use of machine translation in the study of deep neural networks, which spoke of the use of "regression neural networks", which is understood to be the most advanced technology in the field of language recognition.

More powerful "Google Brain"

Quoc Le said in the article, they found that the new method is better than other machine translation algorithms, but this is only a "deep learning" of the application of the future "in-depth learning" technology will also be used for network problems, automatic description and emotional analysis and so on.

In order to take advantage of these advanced algorithms, Google will not be able to expand its "machine neural network" scale, rather than limited to the area of image and speech recognition. The founder of the "deep learning" concept, currently working for Google's Jeff Han Ding Geoff Hinton, said: "Like a pigeon's brain, although it has good eyesight, no one talks to a dove." ”

In fact, even a pigeon with a fairly small brain capacity, its brain computing power can easily surpass the current world's most advanced "Machine neural network" (including "Google Brain"), and in handing to join Google, but also announced in the future to help Google build the world's largest "machine neural network" to "deep learning" Conduct a more comprehensive study.

Translator: Xiao Kai

(Responsible editor: Mengyishan)