Chestnut Li Gen from the concave temple Quotation Report Public Number QbitAI
Quotation Report Public Number QbitAI
Google’s Dropout patent came into effect on June 25.
Valid for 15 years.
The so-called Dropout is a commonly used method for deep learning and training of neural networks. It was proposed by Hinton in 2012 to prevent over-fitting.
Three years ago, because Google applied for a patent for Dropout, it has already caused an industry upset. At that time, there was a hot post on reddit, and the words and enthusiasm of Google, and the Hinton who did not receive the Turing Award at that time were all given to the greetings.
History has proven that it is useless. Today, Dropout is still a must-have weapon in the AI space, and —— Google has definitely obtained a patent license.
Don't panic first.
Look, the netizen @peng321 says:
Don't worry, just wait until 2034, it will expire.
Of course, before the Dropout patent expires, everyone, all the people, companies, and organizations that use Dropout may face an embarrassing situation.
Reddit User @rantana summed up:
What is Dropout?
The name of the patent is: a system and method for solving neural network overfitting (in English).
Take a look at the summary and you'll know the basics of Dropout:
This is a system used to train neural networks. Feature Detectors are connected to a switch, at least some layers of the network.
In each training case, the switch randomly turns off some feature detectors based on a pre-configured probability.
Then, the weight of each training case is normalized to apply the neural network to the test set.
As for why it can prevent overfitting, Hinton Grandpa and his friends, in the CNN fame ImageNet Classification with Deep Convolutional, roughly said this:
Each time you turn off half of the feature detectors, each training is a different network. Then take the average of the various networks for prediction. This can improve the stability of the model, or generalization ability, to prevent over-fitting.
The inventor of the patent document is also the author of this paper, but the applicant is Google:
Regarding what is included in the patent, Google lists20 articles, counted a variety of thoughts and unexpected training steps:
It's a good use note for Dropout.
If you are confused in the process of hitting the fitting, you can go in and check it.
Of course, this is not the point.
The key is, what restrictions will be imposed by using Dropout in the future?
Major obstacles to open source
The netizen named mtanti asked:
That is to say, we can't use Dropout in the future?
The answer downstairs (Nicolas Guacamole) is wonderful:
Whoever used it, take the drone to get rid of it.
The banter is one of the attitudes. Telling it in a normal sentence is:
Google will not really use this patent.
Of course, not everyone looks at it this way.
In fact, as early as three years ago, this patent has been approved and triggered a serious rebound.
Dropout patent application grant information
As for what kind of future is foreseen, this kind of emotion will be born. The netizen (AnonMLResearcher) carefully analyzed:
Someone once said, "Don't blame the players, you have to blame those who make the rules of the game. ”
That is to say, we should not blame Google and should blame the patent system.
sinceAllow such famous abstract concepts to be registered patentsGoogle also has their legal rights.
I am just afraid that this may cause significant damage to the academic research of machine learning.
In the field of vision, SIFT and SURF have been registered patents. In this way, open source libraries like OpenCV are very troublesome.
Repo does not contain the "Nonfree" & rdquo; module, so users have to build from the source code, it is very troublesome.
In the future, the open source machine learning library will still encounter the same thing.
Card neck warning
And the basic algorithm such as Dropout is a Google patent, so the focus is on the open source community.
Netizen (AnonMLResearcher) said:
Such patents can have an impact on any young machine learning company, making it harder for them to attract investment —— now Google has intellectual property rights for many algorithms, and (if using a patent code) may be sued at any time.
"The question of Xu Xiaodi", which was once widely discussed, was once again discussed.
In May of this year, at the Shanghai Academician Salon, many academicians such as Xu Kuangdi, an academician of the Chinese Academy of Engineering, asked sharply: How many mathematicians in China have invested in the basic algorithm research of artificial intelligence?
The status quo is that although the application of Chinese AI is in full swing, the underlying framework and core algorithms of independent intellectual property rights are very scarce, and more rely on open source code and algorithms.
Professor Kong Dexing, director of the Institute of Applied Mathematics at Zhejiang University, who was interviewed by the Science and Technology Daily, called for: If there is a lack of core algorithms, when it comes to critical issues, it will still be called “card neck”.
Professor Kong said that open source code can be used, but professional and targeted, the effect often can not meet the actual requirements of specific tasks. Professor Kong believes that whether or not the core code will determine whether there is a chance of winning in the future AI<Intelligence Competition”.
This year's repeated examples have proved that the core algorithm relies on the lighthouse, and even if word of mouth is like Google, there may be a day of disconnection.
So Dropout is a Google patent, not only about open source, not just research, but also about the more deadly autonomous core algorithm and the "card neck" behind it.
QubitQbitAI · headline signing author