Dialogue with Google scientist Jeff Dean made me start to find that mathematics and statistics are quite interesting.Our dialogue theme is “machine learning”, which stems from Jeff Dien and his team’s famous “Cat Face” experiment.Google uses 16,000 computer processors to build the world’s largest electronic simulation neural network, and shows 10 million videos randomly selected from YouTube to this artificial neural network.As a result, in an environment without external instructions, this artificial neural network actually learned to identify the cat’s face.

You know, Jeff Dien has never told the concept of this artificial neural network “cat”, and has not even provided it with an image marked as a cat.”This seems to be a process of natural acquisition in the growth of newborns. It has self -awareness. It has found what cats look like in the YouTube static picture that has never been marked,” Jeff Dien said, “For example, let’s take care of it.The child is put in the car and asked him to observe the outside world from the window. He will see a variety of motor vehicles and bicycles, and no one will tell him what these objects are. He summarizes it through his own observationsSimilar things “”

This is the difference between machine learning, it produces “self -learning”.

In the vast ocean of artificial intelligence, machine learning is the hottest topic.If you recognize the definition of artificial intelligence in 1950 if you recognize the father of computer science in 1950: If one person cannot judge whether to chat with him or a computerArtificial intelligence; then the new meaning of machine learning is to make machines think like the human brain.

“Machine learning is a discipline about computers based on data construction models and use models to simulate human intelligent activities.” Dr. Li Hang, a machine learning expert and chief scientist at Huawei Naya Ark Laboratory, has defined popular science for machine learning.This is still slightly obscure for laymen. In fact, many details of our daily life have been implanted, such as voice recognition, spam intercepting, unmanned cars, and even Taobao.com.

Take the voice assistant service Siri launched by the iPhone last year, how gentle and considerate personal assistants.She provides mobile phone users with weather forecasting, catering suggestions, awakening inquiry and other life services. It is okay to talk to her in the “The Big Bang Theory”.Under the bombardment of hundreds of questions every day, she is becoming more skilled. Not only will she not answer the fact that she will not ask, but she will even judge whether you are drunk and need to help call a taxi.

The performance of some machine learning systems has even surpassed humans.For example, the Swiss AI laboratory of Lugano University in Switzerland designed a system to defeat human experts in the competition of traffic signs.

The higher -profile display was at the end of October 2012. At an academic conference organized by Microsoft Asia Research Institute, Nankai University, Tianjin University, Microsoft chief scientist Richard F. RashidIn the auditorium, the computer program was identified to his speech content, and these contents were also displayed on the big screen above him in English.

After that, after saying each sentence, he paused, and the computer translated these words into Chinese, and also attached Chinese dubbing very similar to his voice.You know that Rashid has never spoken or speak Chinese. This show has won applause that has won thunder.

The international media did not give up this hotspot. The New York Times published a front page, using the words “really great!” And other words to express the optimism of machine learning prospects; then, “New Yorker” also published an article response”This allows us to move towards the real intelligent era.”

But Jeff Dien told me that his experiments are another level of machine learning, and the technology used in the above examples is at two dimensions.”It’s more appropriate to call it” deep learning “. This is a branch of machine learning.” Li Hang said that this is also a way and way to realize machine learning.This process is roughly the neurons of the computer simulation of the human brain, and then many computers built a neural network similar to the human brain. This is like a small -scale new brain. This “brain” can perceive like the human brain., Identify, memory and even think.

Although several large academic organizations are studying at machine learning, “Google hopes that this technology will not be limited to a specific field, but can make it larger and build a relatively large model, so that the computer will make the computer make the computer make the computer.The system can get a greater understanding from the original data. “Jeff Dien said.It is not known whether these artificial brains can pass the Turing test, but the results are not few.Jeff Dien and the Google Voice recognition team tried to cooperate, with 800 machine training for 5 days, and soon reduced the identification error rate by 25%.The results.At present, this technology has been applied to voice recognition technology to make voice recognition more accurate, which is reflected in the Android 4.1 version of the operating system “Jelly Bean”.

This is a very expensive scientific research project. These artificial neurons are expensive and require a lot of data.Because the larger the amount of data, the more you can participate in the number of computers in the machine learning network. During training, the learning ability of this “artificial brain” will be stronger.”We are actively expanding the system to train a larger -scale model.” Jeff Dien said, “Although there is no recognized way to compare the artificial neural network with the biological brain, in order to let you feel the so -called so -called so -called so -called so -called so -called so -called so -called so -called so -called so -called so -called so -called so -called so -called so -called so -called so -called so -called so -called so -called so -called so -called so -called so -calledThe “larger scale ‘can still make a very rough comparison with the human brain -the ordinary adult brain has about 1 million trillion connections.

Indeed, animal learning is perceptual, and the learning of machines is pure logic.Compared with the human brain with 80 billion neurons cells and 1 million linked, the “artificial brain” is probably much clumsy.Therefore, when I was developing and expanding whether artificial intelligence would develop and expand, and to invade human life, Jeff Dien just laughed. “You know, the current machine learning we call is limited to recognizing the recognition of recognitionKnowing categories can only handle the simple work of cognitive categories, “this quite rigorous scientist explained, and what people really want is to obtain the output that matches the input and automated input, for example,” Today at noon today, “Today at noon todayWhich restaurant to eat “and” how to coordinate the relationship between friends and work “are currently not available for machine learning.