Home > News content

Baidu AI's enclosure movement: what has been done in a year?

via:博客园     time:2018/5/29 15:55:53     readed:720


Wen / Guo Yipu

A year ago, Baidu opened the first developer conference of China's AI.

Baidu founder Robin Li said that the AI era will be the age of developers.

A year is coming, what can the giant shoulder do? What is the appearance of the AI ecology of the Baidu?

Behind this, there may be some changes in the tide of the times.


From the results alone, Baidu has really launched a battle.

Around the developer, from the point, to the line, to the face.

From technology

Technology oriented

be directed against

EasyDL enables small white users to train products of deep learning models according to their needs.


Four step method for creating personalized image model by delta EasyDL

For example, you can use this EasyDL to train a model to identify cats.


Or a dog.


Or do something more practical, such as a convenience store clerk who can automatically examine whether the display is in line with the requirements, without the correct alignment of the drinks, chips, and biscuits, and the factory automation assembly can automatically examine the components of the spare parts with the trained model. All, no manual inspection is needed.


A well trained model can work in different scenarios.

AI entry and practice seem simpler than imagined.

In fact, since the launch of EasyDL, there have been some examples of reducing efficiency and increasing efficiency.

In the Chinese food and drug testing institute, with the help of EasyDL, they build a model of Chinese herbal medicine identification, which is used to assist the researchers to identify the drugs. At the same time, it is convenient for the drug supervision bureau to inspect the law enforcement personnel to quickly identify the drugs and improve the efficiency. Up to now, the model has realized the recognition of 42 common Chinese herbs, the accuracy rate is more than 97.4%, the training time is less than 10 minutes, and the speed of single map recognition is less than 50 milliseconds.

In foreign countries, the American retail security provider Checkpoint, based on EasyDL, has built an intelligent grab machine to identify whether there are unpaid goods in the lower layer of the supermarket shopping cart. The accuracy rate of the model is over 95%, far beyond the accuracy of sensor judgment.

And some.

Engineer oriented: definition and evaluation system

Of course, opening up EasyDL to novice users is just to cultivate more flowers and plants for this ecology, solid soil, and more central.

In this circle, Baidu's tactics are: issuing guidelines, drawing routes, training, building circles, and assisting more engineers to transform AI engineers.

Like publishing.

Baidu also designed a

In this.

The direction of engineers' upgrading is still lacking, and there are still some specific learning methods and opportunities.

So Baidu also offered a deep learning and education system and a PaddlePaddle AI contest.

In April this year, Baidu launched a PaddlePaddle open class to provide a rich depth of learning curriculum resources, online to help developers master the PaddlePaddle deep learning framework from scratch. This is also the first set of the industry's deep learning education alliance partners to complete the PaddlePaddle in-depth learning Chinese video course including theory and practice.


And for domestic developers, PaddlePaddle, as the only open source in-depth learning framework in China, has advantages in performance and technical features, and is also considered as the only open source in-depth learning framework.

On the one hand, PaddlePaddle integrates many kinds of neural network and deep learning algorithm such as CNN and RNN, and supports many kinds of hardware such as CPU, GPU, FPGA and so on.

On the other hand, as a parallel and distributed deep learning platform, PaddlePaddle supports multi machine multi card parallel, data parallel, model parallel and so on, and also supports the deployment of privatization.

In addition, in the way of invocation, the training part of PaddlePaddle supports browser, client and other ways of calling.

More importantly, PaddlePaddle can provide developers with comprehensive first-hand Chinese documents, greatly reducing the entry threshold for developers.

So after full open source in 2016, PaddlePaddle took the lead in the domestic upgrading of traditional industry, showing that the strong industrial application capacity has now opened nearly 20 industrial models, including click rate prediction, speech recognition, text classification, image classification, Machine Translation, object recognition and so on. In the internal and external products.

Enterprise oriented: prairie fire and acceleration

Even at the expense of free, low discount open interface, reduce access threshold.

Since last year, Baidu has announced a free face recognition interface and a full range of voice interface policy. At this point, Baidu has been free to open the most commonly used AI technology from perception to cognition, providing full free AI capability support for developers and partners.

And Natural Language Processing basic technology series interface, the user each interface has 5 QPS default, if the need for higher QPS resources, developers can apply on demand, audit pass can continue to call free, up to up to 100 QPS per interface.

Understanding and interactive technology UNIT, users can enjoy free of charge 3 Bot (account), each Bot 40 QPS permanent free quotas, if special needs, users can apply on the Baidu AI open platform official network, professional customer service will provide 1 to 1 support services. This is also the largest semantic technology capability open to the public at present.

Yu Youping, general manager of Baidu AI technology and Ecology Department, recently presented a set of data. He said that Baidu AI is basically billing in the way of QPS, QPS represents the number of requests per second or per second request concurrency, and 1QPS is equivalent to 86400 calls a day. Take the face interface as an example, Baidu provides 10QPS with free interface for each authentication enterprise, which is equivalent to an enterprise's free calling of more than 860 thousand times a day.


At the same time, it should not be overlooked that this is also a developer support policy that Baidu had never seen before.

The Chinese giant hopes to create an open, equal and caring AI technology ecosystem through a series of open strategic means.

And there are some small goals:

Among them, the results of the accelerator have been shown.

Last October, the first Baidu AI accelerator was launched at the first station of the AI developer's real battle camp, then after more than two months of registration and evaluation, through multiple rounds of data screening and interviewing from the last thousand projects, the top quality enterprise was finally selected to enter the first AI accelerator.

In the first Baidu AI accelerator, the member enterprises covered 17 big industries, including smart home, intelligent agriculture, intelligent medical treatment, intelligent justice, intelligent customer service, AI chip, automobile service and many other fields. More than 90% enterprises succeeded in access to Baidu AI technology, service over 30 million enterprise customers and hundreds of millions of people Users enjoy the AI service.

During the entry of Baidu AI accelerator, 7 partners, such as visual technology, human intelligence and CELLA, issued 9 new products with Baidu AI technology; 15 partners such as Tian Zheng polymerization, intelligent point, I bought net and so on. With the help of Baidu AI technology, the products and services were greatly upgraded.

At present, the second phase has been opened. The selected enterprises are involved in service robot, chip, computer vision, voice, new retail, human-computer interaction, cloud platform, VR/AR, preschool education, animation, intelligent tourism, image processing, online data application platform, human resources, hotel intelligence and other fields.

Technology output to build ecology

In the past, Baidu, which started with search and sat firmly in the traffic flow, owned the alliance and distributed the traffic.

However, the AI curtain opened, and Baidu's combined boxing turned into technology, investment and ecology.

How to build ecological barriers through technology export?

First stage accelerator membe

The car master's self-definition is

But using AI technology, the car owner can issue a car problem by voice, text or picture, and the efficiency of the AI system will be different on the basis of speech recognition, image recognition and semantic understanding.

Behind the technology support, it's from Baidu. Through the Baidu AI accelerator, the Baidu AI accelerators access the Baidu understanding and interactive technology (man-machine dialogue customization platform), combined with its own rich knowledge base in the auto field, and successfully created the intelligent question answering system of the automotive domain by using the two skills of Q & A and multi wheel dialogue on the UNIT platform. Not only did the response speed increase by three times, but also the accuracy of the answer reached 95%, which greatly improved the user experience.

In addition, such a project is still connected to PaddlePaddle deep learning technology, and achieves the improvement of matching accuracy.

Using technology to open up, more businesses embrace AII, are passing through

The times are so different that the sky is changing and the Tao is changing.

China IT News APP

Download China IT News APP

Please rate this news

The average score will be displayed after you score.

Post comment

Do not see clearly? Click for a new code.

User comments