when using Google company can build a 16000 processor by watching Youtube video to properly identify the cat face “virtual” brain, it marked the artificial intelligence (AI) technology is an important turning point.
this emerging AI algorithm of computer data, and is often referred to as “deep learning”. So-called “Google” brain than ever are said to be twice as high precision digital image recognition system, the Internet giant Google on large data processing of another victory.
in fact, as early as in 2012, the New York times report said: “this study represents a can be used to reduce the operation cost of a new generation of computer science and computer and the feasibility of processing data was obtained from the large data centers, and bring huge development for different fields, such as machine recognition and perception, speech recognition and language translation.”
in the past two years, Microsoft released a use of deep learning in Skype instant translation between two languages. Facebook hired this area leading experts to improve their image recognition, and other service tools, from Twitter to Yahoo, the big Internet companies snapped up startup involving deep learning.
but in the technological change, a named Alex Krizhevsky researchers show that to take advantage of unique “self learning ability of artificial intelligence to analyze digital data, don’t need a lot of computer data do support. A paper published in the same year also have the same view, the author took a computer, at least in a particular image recognition test is better than that of Google company with 16000 processors.
this is a very expensive computer, large capacity memory, the top of the GPU and equipped with its own dedicated computer chips have multiple computers in combination with the effect of the operation. But it is just a computer, it shows that you don’t like Google have computer cluster can also use the ability of deep learning.
using the artificial intelligence technology still needs certain professional knowledge, this is why the Internet giant at cultivating talents. It is because they have large data centers and substantial financial resources, companies like Google are to bring the technology to the world. At present there are still a lot of data scientists using a computer can only use deep learning algorithm to solve the problem of their own, the computer may be used by ordinary consumers in this game.
in Kaggle websites, data scientists on behalf of other companies and organizations to solve problems, deep learning is one of the preferred tools. According to Kaggle chief scientist Ben Hamner, single computer has can be used to solve everything, including image analysis, speech recognition, and even chemical informatics.
researchers at Stanford university, says Richard Socher deepens the use of deep learning, using a computer can achieve the purpose of the recognition of natural language, this indicates that the artificial intelligence technology can bring benefits to some small companies. Socher said the processing system is very easy to deploy, anyone can do it.
at the same time, some startup companies began to build the cloud services to provide deep learning tools, and other companies to launch of the software other than the Internet giant and consulting services, to some extent this is conducive to technology diffusion. After all, the world only then several companies of computer cluster can reach the scale of Google, Facebook and yahoo, small ordinary companies can also use a single computer to carry on the depth study work.
the GPU is the abbreviation of the image processing unit, these chips was originally intended to quickly generate on behalf of the games and other designed for the application of highly visual images, but because they have a certain math ability, also can play a variety of tasks. In fact, these tasks include deep learning.
deep learning trying to mimic the behavior of neural networks in the human brain, nature created a multi-level software system, once the configuration is correct, when they analyze the data more and more long, can realize self learning. The traditional machine learning need manual operation in a wide range of human engineer, and deep learning is not needed.
the multi-layered neural network parallel operation involves many computer chips, like Google company 16000 processor parallel processing, but you can also through the GPU to realize the parallel processing, the top of the GPU can contain more than 2000 processors.
Alex Krizhevsky use machines equipped with two GPU run deep learning algorithms, treatment effect than Google is equipped with 16000 processors, basic CPU computer better. Is different from the large amount of data to support Google his ingenious use of the operation of the algorithm and the GPU close operation, do not need to sent a large number of data on the network can also be in-depth study.
now Alex Krizhevsky at Google, his deep learning a startup company recently acquired by Google. Is the same as other Internet giant Google is exploring how to using the GPU in his deep learning work, this is mainly for the use of the GPU for deep learning opened up a shortcut.
in Kaggle data the scientists used a $3000 worth of single video card game this run deep learning algorithm, to solve the problems of the image and voice recognition, and the technology also has great use in other areas. Kaggle won the first round of the competition, it has developed a deep learning machine, it can be predicted based on the chemical structures of some molecular biological reaction, using only a single system, this process using the previous image and video processing technology.
of course, to say the least, allocation of 16000 processor system is sometimes more useful. Such as Google and Facebook company with this system can quickly analyze large amounts of digital sound, image and realize self learning. But if your data set is small, a single system can still provide a certain standard of artificial intelligence, this is one of the traditional learning system cannot reach.
according to Socher, deep learning involves two stages of the operation. First is the learning phase, the system by analyzing the data of self training, familiar with the corresponding operation. Arrived after the second stage, the stage is through the system of artificial operation to deal with the problem. Learning phase requires a computer to have high processing ability, but in many cases, a single processor can meet the requirements, it depends on your need for speed.
annoying is that the Internet giants to buy all the talent in the field of deep learning, these talents are still very important to build a virtual neural network, the depth of the neural network training is not only a science, or an art, many parameters are often used to train neural network to rely on intuition.
in other words, a lot of deep learning algorithm is open source, this means that anyone can take advantage of them, and start-up companies. A startup called the Skymind including San Francisco, the company is committed to cultivate good at complex algorithm data scientists. At present, large companies such as Google and Facebook leading the revolution of artificial intelligence, and the future will be more and more people join this revolution.