How Baidu just Paddle platform for open source

Tsung-jen, Huang Xin co-editor

Paddle depth, Baidu today open source learning platform has raised very much developer interest in the field of artificial intelligence, including on-the-job training in Tensorflow and Caffe until some developers. But given the deep learning open source platform is not much, masses are as keen to eat melon as a development priority, just want to know–how about this platform? What others think of this platform? And this platform different from Tensorflow as well as Caffe?

How do ▎ the platform itself

Have existed before the Paddle itself in open source, beginning in 2013, when Baidu depth depth laboratory realized in neural network training, along with computational advertising, text, images, voice training, such as the rapid growth of data, traditional training platform based on single-GPU has been unable to meet demand, led by Xu Wei, the lab built Paddle (Parallel Asynchronous Distributed Deep Learning) how parallel GPU this training platform.

But today, open source simple model of the Paddle is not 3 years ago, Paddle 3 years ago may also be a deep learning platform-independent, not well supported from other platforms, data access needs. But today’s Paddle is already stressed, it features a Spark with the PADDLE is coupled, a depth of heterogeneous distributed system based on Spark. And after and Baidu-related business of “tight friction”, it has iteration two versions: from the Spark on Paddle schema version 1.0, to Spark on PADDLE schema version 2.0. According to the platform open source rules, presumably in Baidu is very handy for internal repairs after a series of bug lab was finally going to Spark on PADDLE and heterogeneous computing platforms are open source. As to why Baidu to raise revenue, and the reason we all know

Deep learning platform, there are many bug–to attract more developers to try and use advanced learning technology, certainly help to improve the grade of Paddle.

▎ Evaluation of outsiders on the platform        

Know Jia Qingyang answer, is now more positive evaluations.

1. very high quality GPU code

2. very good RNN design

3. the design is clean, not too much abstraction, it is much better than TensorFlow. Juicy Couture

4. high speed RDMA part seems to be no open source (possibly because the RDMA for cluster design has certain requirements): Paddle/RDMANetwork.h at master · baidu/Paddle · GitHub

5. design thinking more like first generation DL framework, but the paddle has been for years, also has historical reasons.

5.1 config is hard-code protobuf message, the extension could have an influence.

5.2, you can see many interesting designs similar to the legacy: using the STREAM_DEFAULT macro, and then directed to a non-default TLS way stream:Paddle/hl_base.h at 4fe7d833cf0dd952bfa8af8d5d7772bbcd552c58 · baidu/Paddle · GitHub (Paddle off-the-shelf Mac are not supported? ) Juicy iPhone 5 case

5.3 in the gradient using the traditional coarse-grain forward/backward design (similar to the Caffe). One might say “so paddle does not auto gradient generation”, this is not the way, autograd existence regardless of the size weight and op. In fact, the TensorFlow aware of fine-grained operator super slow speed, gradually back to coarse-grain on the operator. Currently only see here. All in all was a very solid framework, Baidu development skills are good.

Juicy iPhone 5 case

Estimated evaluation of many people read Jia Qingyang, one Baidu data before we post the following engineers perspective CTO Ying Ning’s evaluation

Looked at from design concepts and Caffe is like, but the network model is not as easy to define as Caffe. Maximum contribution is made, distributed, improves the model’s speed. Details will have to look at the code and get started again.

Another with the above two opinions contrast a deep learning scholars

 Tensorflow schema can be thought of as an upgraded version of theano, theano years earlier than the Caffe, but Caffe first train out, and released a successful of convolutional neural network model to get more attention. Relationship between Tensorflow and Caffe fine much, could learn from the Caffe some implementation techniques, essentially nothing. Baidu is likely to be seen after successful implementation of Caffe, much imitated Caffe, while trying to change some things to make it look different from Caffe.

I assume Caffe who directed it, who uses other tools (tensorflow, keras, theano, torch,mxnet) are not into it, say for a few days and then … … Github attention look at it in a month and the amount of code on github to find someone else to write what you know whether he could spray (later no one can see him attend kaggle or other games or in scientific publishing). Now each company released its own deep learning framework (or machine learning framework), such as Microsoft, Amazon, Yahoo, seems to have been no major movement.

▎ Different from the platform with Tensorflow and Caffe

Lei feng’s network (search for “Lei feng’s network”, public interest) applied for a Paddle public beta today, are also being reviewed, although not direct download experience, but the difference between the other two platforms and is not without a trace. Our Caffe before, Tensorflow understand, as well as Paddle data released today.

How Baidu just Paddle platform for open source?

Interfaces audio

Caffe ——cmd, matlab, python

Tensorflow——python, c++

Paddle ——python, c++

(Note: Python is the language developers primarily use only Caffe changes to model internal to use c++. (If there is dissent, welcomes developers to further exchange)

In General

1) Caffe can be said to be the first industrial grade deep learning tool, was founded at the end of 2013 prepared by UC Berkely Jia Yangqing development language a CNN implementation of excellent features, Caffe in computer vision area is still one of the most popular Kit.

Caffe development language C++ and Cuda, very fast, but due to some remnants of historic architecture, its flexibility is not strong enough. And support for recurrent neural networks and language modeling is very poor. Caffe supports all major development system, use a moderate difficulty level.

2) Tensorflow is a second generation Google open source advanced learning technology, RNN API implementation is an ideal, it uses the symbol method for vector operations, develop speed quickly.

Tensorflow better system only supported by various Linux systems and OSX, but its support for the language is fairly comprehensive, and includes Python, C++ and Cuda, developers document is not as comprehensive as Caffe, so use more difficult.

3) and this time Baidu’s Paddle, as the depth of heterogeneous distributed system based on Spark, through the use of GPU and FPGA heterogeneous computing upgrade data-processing capacity per machine, temporarily access to the industry’s “fairly simple, clean design, stable, fast, occupying smaller video memory. “Evaluation of it by using the GPU and FPGA each machine data processing and heterogeneous computing capability has important contacts. But how, still need to wait a few days to observe people experience.


Read Google Pixel and Nexus as well as AI First

Lei feng’s network (search for “Lei feng’s network” public attention) by writer Yu Di, Vice President of China’s cultural industry investment fund.

Inadvertently sees Google friend to share intelligence, and from there, talk about the things Google related.

Read: Google Pixel, and Nexus, as well as AI First

“Really Blue” (true blue)

Reflects the appearance of Google technology, too little hardware products technology around gene and the cynical geek style (ridiculed rival iPhone generations of color design and marketing too Orthodox taolu); Google employees on October 7 to get volume production purchase Pixel XL, reflecting the new phone is now stocked in 6 countries.

Read: Google Pixel, and Nexus, as well as AI First

2. Pixel cooperation operator is United States top two operators Verizon and cooperation rather than the Nexus series before after higher-ranked T-mobile.

Google for Pixel can be seen more attention, also Google has recognized the 6-year life cycle of the Nexus product line sales downturn bound is one of the main reasons why carriers competitive relatively weak (contract of purchase operator in North America is the main channel for mobile phone sales).

Read: Google Pixel, and Nexus, as well as AI First

3. the mobile Google’s own brand “g”, rather than the Nexus series logo policy before (Nexus of each model are exposed joint ODM vendors to brand).

While can see Google for intelligent phone of industrial design began itself led, and production manufacturers of cooperation mode has by zhiqian of ODM to OEM (Pixel by HTC for OEM generation workers), on the can see Google build independent brand phone of beginning and determination, and willing to for bear brand cognitive risk (Google phone user also main limited in very guest, and development groups, and Google fans groups). Disney iPad Mini Case

Well, have a bit of fun, warm up, the following talking points.

| Information relating to the Nexus interpretation

Introduction Nexus phones are compromise

Google in the past ten years, adhering to the strategy of Mobile First and tried to buy MOTO to produce own-brand mobile phones, but is subject to several factors which led to integration had not been successful, eventually had to retain some MOTO after MOTO patent sold to Lenovo. One of factors including genes are different: Google is the Internet, MOTO-brand manufacturing; and Google’s own-brand mobile phone broke the Samsung Android camps such as the interests of the other partners, and so on.

Zhihou, 2010 began, Google through and ODM Manufacturers (HTC, and LG, and Samsung,) launched joint brand of strategy for phone development, and design, and production and sales, Nexus new model launched of frequency Basic is annual 1 to 2 Biggie (each Biggie including several different configuration model of small paragraph), Google for Nexus of expects is for other Android phone manufacturers provides a most frontier of hardware and software model standard, So as to help promote the Android Smartphone of the camp layout.

Global sales of Nexus

While Google does not disclose the specific handset shipments, but from the Google internal staff, external media or advisory bodies to the information gathering here, each Nexus mobile phone shipments are less than 5 million, some models such as Nexus 10 overall shipments of less than 1 million;

Why Nexus low sales?

(1) Google the location of Nexus are demonstration models, does not expect Nexus to be consumer-oriented selling cell phones and Google in revenue (for example, in fiscal year 2015), hardware revenue less than 1%;

disney ipad

(2) subsection (1), the Google Nexus resources are more limited, including the selection of operators (more cooperation is ranked before T-Mobile, restricted channels), ODM Manufacturers willingness to promote the Nexus of weak, Google team on supply chain management and capacity are relatively short, and so on;

(3) Google services framework is limited in China, leading native Android system in China popular, Nexus software as standard and the service cannot be used in mainland China (except via VPN over the wall), leading to further cut in the size of the audience;

(4) the public Google smartphones lack of brand recognition, even in North America, consumer hardware products, especially for Google Smartphone, awareness and acceptance are extremely limited.

 Nexus have a negative impact on Google

(1) brand series of chaos: the frequent replacement of the joint ODM brands, product lines naming frequent adjustment;

(2) unable to get specific classes of mobile user information and missed part of the mobile Internet data entry: user using brand manufacturers such as HTC’s own application market, third-party photo backup, cloud services;

(3) impede the layout and development of Intelligent Home: smart home brand products a few Nest of Google acquisitions, IOT ecology has not been established;

(4) the hardware product line gap: Google has a laptop before Chromebook Pixel, Tablet PC hardware products such as Pixel c, the missing Pixel own-brand Smartphone (AR-enabled phones to work with Lenovo Phab 2 Pro-Project Tango bounced to the end of the year sale, module cell phone Project Ara no below);

(5) restricted AI strategy development: as the entrance, scenes, such as the lack of access to information, resulting in data integrity be affected further influenced AI research and development, such as machine learning model, and so on.

Read: Google Pixel, and Nexus, as well as AI First

| AI strategy and the First Pixel on the Google Smartphone

 Google in the field of AI development

(1) development of AI of several major elements: big data, algorithms, computational capacity/resources;

(2) for Google, and Google’s team of top engineers and scientists based on the algorithm model of development and continuous optimization is a strong protection; servers, Super-ICC and other hardware resources will not be a problem for Google then this link leaves the data, namely AI Giants all over the world face the core issue;

(3) data: get orders of magnitude more data (get tag after processing), more data dimensions, data, the broad range, after training for input into the model, model output will be the more accurate intelligence.

So, Google is currently developing several strategies of AI is relatively easy to read, that is–to expand intelligent terminals (including hardware and software), open Google Assistant platform/ecosystem (Google Assistant is one of the Google AI interactive interfaces to be rendered), competition for competitors to obtain data and information portal.

Google hardware product development ideas

(1) official with the hardware product

Although hardware made in the area of achievement or showing determination not recognized by Google several initiatives this year, you can see it was decided to formally with the hardware product.

For example, Google appointed former President of MOTO as Google hardware product General Manager (make up for the lack of Google’s hardware, supply chain management, such as genes or ability), recently has invested at least $ 3 million Pixel mobile phone marketing and advertising (in the same period more than Apple and Samsung) and follow-up will be investing hundreds of millions of dollars in advertising (according to Reuters forecasts), Chromecast products such as hardware upgrades and published (often abortive hardware product development before the change), and so on.

(2) hardware products in order to develop AI

Apple in 2015 of 4th quarter earnings, only 68% per cent of iPhone revenue, if you count the other hardware products, you can see Apple’s main source of income is the hardware;

Google hardware products are not expected to contribute much revenue for hardware positioning, Google hopes to continue to create demonstrations and official standards on the one hand, such as the release of Daydream View mobile phone boxes, is to unify the jagged mess on the Android platform product standards;

On the, more important of, is above 1th by mentioned, Google hope through hardware products to gets more dimension degrees, and more large level, and more continued of time cycle within of Internet (containing mobile Internet), and real networking (this is Google this launched Google Home intelligent speaker of reasons one of) of big data, in gets data of while, through Internet + real networking of equipment to user provides diversification, and platform/ecological type, and continued optimization of AI service, To further stimulate the user provides data entry, form a positive feedback, ensuring the rapid development of Google AI strategies, or else Google AI products and service level will stagnate in the early days of Sundar said.

(3) add that still just opinion, Google in addition to AI forward-looking and a leading position in the field, in terms of hardware products and services, just followers of roles played. disney ipad

For example, Google Home reference to the Amazon Echo, before the Chromecast/Google WiFi products such as domestic manufacturers already have the functionality of similar products, Pixel appearance/price iPhone7 remain consistent as far as possible, Pixel brand of unique feature has been implemented in other Android makers, and so on. However, this is because the hardware simply to implement paragraph (2) goals, not Google’s core competitiveness.

About Pixel cell phone

(1) the conflict of interest

Google brand phone again will result in that year to acquire MOTO Android camp that is thrown when the other manufacturers of conflict of interest, however, Google has made it very clear, Mobile strategy has shifted to the AI of the First First, meaning Google complaints about the Android Titans no longer care so much because the next Google Assistant eco-building maintenance of the weight will be larger than the Android ecosystem.

(2) the Pixel’s contribution to AI

As mentioned above, here a few examples–Pixel provides more camera functions, as well as the Google Photo unlimited cloud based backup, additional system user photos feedback (for example MIUI baby photo album features manual adjustment function, essence, the data labels for the picture calibration), provides data for the Google of image recognition model input.

Google Image Assistant provides OCR recognition based interactions or voice interaction, data entry for Google’s speech recognition model users in mobile phone operating system will feedback to Google accumulate large data segments, and Google Account after the match, as user data part of the supplementary, and more precisely within the Google Home Internet environment, such as the AI service.

(3) promote resistance

Is the above mentioned former President of MOTO at the helm, still continues to invest in huge advertising costs, cannot quickly compensate for the Pixel in the near future a couple of obvious weaknesses:

 Google phone brand recognition outside developers, geeks/fans/audience development;

Barriers to market entry;

The iPhone7, high-end prices (lower the Pixel value in the Android camp);

Relative to other marketing channels are limited;

In addition to the built-in Google Assistant, support from Daydream VR, Pixel failure to highlight bright spots;

Out of joint brand, supply chain security (marketing, after-sales service), and so on; it is based on these potential obstacles, cannot currently give accurate predictions about Pixel sales.

In short, Google’s strategy and plan the path is quite clear, in all kinds of scenarios of the future Gets a loyal user, advertising revenue is ripe. But this leads to another topic, namely “how Google works on original advertising strategy”, after there is time to talk about.

While sharing so much humble opinion, welcomes the heroes who disregard this communication.

Note: the first figure from

Facial recognition to new heights no face recognition

What is an important foundation for face recognition? Is probably clear mug shots? No, no, no, before Germany marks at the Max Planck Institute of new research results might change your mind. Jeremy Scott iPad Mini Case

Jeremy Scott iPad Mini Case

The Institute’s staff said, they created a whole new face recognition algorithms, called Faceless Recognition System (FRS), that is, no face recognition. In other words, the system can be identified by a fuzzy photo of individual. Jeremy Scott iPad Mini Case

Facial recognition to new heights: no face recognition!

First, researchers will use a clear picture for learning the system, FRS will face in the photo feature, head lengths and other parameters were analyzed. Then if you blur the picture, it has a 69.6% probability would recognize the character. And if there are enough photos for FRS, its success rate can reach 91.5%.

Facial recognition to new heights: no face recognition!
Facial recognition to new heights: no face recognition!

But if the face is covered by a black box, then the recognition rate will be reduced to 14.7% of the machine, but the average success rate was three times higher than the human. Allegedly, this technology can help the authorities identify criminals from huge photos in the database.

Hard to prevent With WiFi signal records will be able to capture keystrokes

Hard to prevent! With WiFi signal records will be able to capture keystrokes to steal passwords

There is a secret security measures called “keystroke loggers”, by recording the keystrokes used to form the password protection of the operator. This is an upgrade in safety protection measure, this biometric has now greatly compromised.

According to foreign media reports, from the United States State of Michigan State University researchers at the University of Nanjing, China, and such method is found, through a common router using a Wi-Fi signal detection to attack key records.

Researchers point out that, in an environment of minimal interference, the attacker can interrupt router WiFi signals to detect users on your keyboard hit record, and then use these data to steal his passwords, accuracy of 77% and 97.5%. Besides keystroke recording in addition to Wi-Fi signals can also be used to read the user’s hand gestures, lip movements.

Operations, in order to collect WiFi signals of small changes, WiKey laboratory researchers using a MIMO Router (multiple-input-multiple-output) functions to carry out the attack operation. Collect WiFi signals and scan the room, with the help of information, researchers can create a map of the indoor environment. X-DORIA iWatch Case

When a person is standing in front of the notebook and is ready to enter, WiKey can grab by the users hands, fingers and the keyboard WiFi signal interference caused by small changes in the data. Researchers explained that: “each of the keystrokes, users and moves the hands and fingers in a unique way, so get in the channel state information (CSI) time series of values to form a unique pattern, we call it CSI waveform of this link. ”

Research team says, get after the keystroke logging data, they developed an algorithm that converts the user’s keystrokes can be recorded to enter text records. Researchers revealed that 77% to 97.5% accuracy can allow an attacker to know the user’s password three-fourths.

However, this method is prone to bug, when a room when there is more than 1 laptop, it is unable to determine the target. X-DORIA iWatch cover

X-DORIA iWatch cover

Micro OS do set off the second revolution of HTML 5

Recently, the app is being tested, the news spread like wildfire, silence, micro application letter for a long time again a cause for concern. Industry analysis on the official test application, mainly from the series appeared in the letter new H5 promotion page to @ Dragon old thief has issued a document that H5 page changes.

Existing ordinary H5 pages, pictures from the ZTalk@ Dragon old thief


Recent new H5 page, and picture transfer from ZTalk@ Green Dragon old thief

As you can see in the figure, slightly new H5 page letter head has been completely concealed, visually very similar to native App splash page. Meanwhile, new H5 switching experience and click the app is similar to switch very smooth.

In fact, the H5 is always trying to remove “can only do marketing” hats. In April this year, micro browser upgrade to kernel X5 Blink then, app support H5 standard, comparable to H5 in the app has been able to build a native experience. In addition, some severe H5 gaming experience and performance is almost the same with original green tour.

Micro application since it was exposed, since always linked with H5. Queiroz believes that emerging applications, will turn the app into a “operating system”, a Web application in “micro OS” to run on, and take the most important step in the commercialization of micro. Vera Bradley iPhone 6 cases

In application form will be used in micro, micro OS most likely to detonate “H5 second revolution” (beginning in 2012 H5 uprising in 2013, failed for the first time), which has formed a subversive impact on mobile Internet ecosystem.

| Flow + communication of the Kingdom of gene H5 ‘s second revolution will change the way mobile applications promotion

Under the very worst mobile Internet ecosystem in the present competition, promotion and talk is the eternal pain, regular promotion, under the high cost and conversion rate, resulting in a large number of mobile applications at risk being stopped to paint the top.

If H5 application included in the micro-OS, thanks to the micro-social property, early admission mobile application vendors will receive the promotion channels and media resources. “A key concern” to convert the user, product content can also be in your circle of friends open socialization, this picture is so beautiful. You know, the app has more than 650 million monthly active users, is a natural flow of kingdoms.

Meanwhile, H5 technologies themselves have a good spread of genes, is well suited to marketing and promotion, this is a native application that cannot be matched by any account “genetic advantage.” If micro-flow advantages and H5 to combine the advantages of, and its scale effect will completely reverse the existing promotion system.

| Talking about experience, the second revolution H5 got up from where there was a fall

It is well known that native apps and H5 application development model is not the same. H5 applications development advocate quick wins, light, low cost, short cycle. Meanwhile, H5 criticism focused on the bad page UI and flow, can only do marketing pages and games, all, compared with native applications, experience too badly.

Back then, H5 from the rush to become synonymous with marketing and game, the Middle has experienced incredibly rapid expansion. The whole process is like a computer suddenly appear in ordinary people’s lives, you would think that this stuff is really good! Both draw and playback audio, can play the game. But most are easy to play audio, so that everyone is rushing to use the computer to go to the movies, not long before computers became synonymous with video media player, they say that, the computer just like that, watching movies or going to the cinema!

We described above in the “computer” for H5, one problem, H5 ‘s potential has not been fully exploited, and industry at an early stage to the most likely to do, and most easily liquidate marketing pages and games.

Back to H5 experience problems. Indeed, early mild H5 feature brings user experience cannot be compared with the native apps. But now the H5 technologies are evolving, the ability to meet user experience has been quite different, especially part of the H5 technology providers have developed to meet the needs more depth of H5 development platform and higher performance engines (including 3D serious game development), only from experience, H5 can be comparable with the original application.

We should be grateful for in an environment of bad-mouthing people who adhere to the H5, they support didn’t let myself die, foreshadowing the H5 the cornerstone of the second revolution, also just H5 got up from the location of the fall.

At this point, we try to rehash the H5 edge, close to the people’s development costs and cycles, better cross-platform capabilities, based on cloud technology, does not account for memory storage and instant updates, and so on.

| Eat a pot of rice, H5 second revolution will break the application of enterprise application information

Specialized enterprise applications to carry out here, because from the global development trend of mobile Internet, mobile Internet next To b, in the field of business services.

The past two years, Enterprise Services subjected to wild growth, the verticals are a representative set of mobile applications, but caused by fragmented information isolated island problems are quickly highlighted. B differs from the c-terminal end users, enterprises, from production to management to sales, integration of information flow is required. How to open up the information flow between the various enterprise applications? It is almost all enterprise services vendors have been discussing the issue at this stage.

“Open” is the birthright of H5 label, it helps to make all kinds of industrial applications connect through other ecological, shared user data and information flow within the ecosystem as a whole.

Meanwhile, all kinds of enterprise applications is applied by collecting, like used to scatter scattered gathered into a large room, eating with a large pot, and comply with the provisions of the same administrator, will significantly reduce the differences between each other, and to establish the absolute Foundation of fluent communication.

So, when compared with the native App in open information flow problems, based on micro-OS H5 application will solve a large part of “islands of information” problem. Vera Bradley iPhone 6 cases

| Complexity into simplicity, H5 ‘s second revolution will be reconstruction model of mobile application development

For now, except by the ReactNative development package of applications, most native App needs both Android and iOS are two different development environments, two independent development, time-consuming, start price hundreds of thousands of, and compatibility is uneven.

Under the new H5 development environment, H5 technologies even allows businesses and individuals just through a simple drag and drop operation, without writing code, will be able to produce similar to today’s headlines, Mito this amazing mobile applications (not marketing pages and little games), and can be used across platforms. By contrast, if the enterprise to develop native applications for the same function, whether proprietary or is outsourcing its development costs hundreds of thousands of, and development cycles range from February-March to half a year. Meanwhile, H5 growing trend of templates allows users of the various industry segments, both for UI or the future will be templates that correspond to basic services can be found. Such templates can be seamlessly enters the application number, even for free. Such changes, said the app OS+H5 bring about a revolution in the mobile Internet market as a whole, is no exaggeration.

Although micro-application number has not yet been formally unveiled, but can be sure of is that “micro-OS” will give maximum value of H5 release or commercialization of micro is the most difficult piece of the puzzle, this with Tencent “connected” position in one continuous line.

Lei feng’s network (search for “Lei feng’s network”, public interest) Note: please contact our authorized, and keep the source and author, no deletion of content.

Vera Bradley iPhone 6 cases

The problem what is the robots we come to hear entrepreneurs say

The problem what is the robots, we come to hear entrepreneurs say

“Robot (Robot) is a machine that automates work device. It can accept the human command and arranged in advance program is run, it can also be based on the principle of artificial intelligence techniques to develop the programme of action. Its task is to assist or replace human work works, such as manufacturing, construction, or dangerous work. ”

Above is the definition from Wikipedia for robots. Believe you are still confused, can help people work is a robot do?

If you want to trace back history of robots, it was even before is as early as 1400 BC, the Babylonians invented the water clock, which is a flow metering time timer, it is considered to be one of the earliest mechanical devices. Later there have been Knights of Leonardo da Vinci, Vaucanson’s duck, and Kempelen Turkey robots and other automated equipment, initially they are known as automata (automatons) or to control machines (self-operating machines) until 1921, when Czech playwright Karel • Capek’s Russell Sam universal robot makes “robots (robot)” popularity of the term.

Now on the market a wide range of robots, industrial, and service; some people also inhuman; physical or virtual. Because has not been framed robots might be an image, it can be said that 1000 people 1000 robots in the eyes of the image, so what are robots this seemingly simple question has become extremely complicated. To this end, Lei Feng network robot-related entrepreneurs have been invited to talk about the issue.

Lei feng’s network (search for “Lei feng’s network” public attention): what is a robot? OtterBox iPad Mini Case

Read technology CEO Zhao Jinglei: abroad there is a term called robotics, it is the concept of the robot, based on this concept, I think it can be defined as some kind of human perception and cognition of the machine. It may be in various forms, is not necessarily human.

Cloud motion technology CEO Tao: about the robot, States and focus are not the same. Intelligent machines and human distinction, extend human capabilities and people not terminated. Now, robots and artificial intelligence is still a new discipline, we can do at the moment is: for a robot to help relieve us from simple, repetitive tasks to deal with anything more needs imagination, creativity.

SI-Chi CEO Gao shixing: definition of a robot very much, if simple understanding is the machine people, giving machines the ability to empower people, called robots. Of course, there are many kinds of human capacity, as now many of these intelligent hardware coupled with perceptions of module module, also known as robots, further combined with cognitive abilities, further in the future and emotional capacity, coupled with the logical thinking reasoning ability in the future. So, plus gives the machine the human capacity of the machine, called robots.

CEO Misa Rokid robot: defining the core of the robot is autonomous machines or assist in the completion of related tasks, so you don’t have to care too much about form.

Pay attention to two points: 1, whether a certain degree completed independently of the outside world’s perception or understanding whether 2, according to the perception on the outside world to understand feedback. This can make it, that is.

SI-LAN technology CEO Chen Shikai: my answer is to actively interact with the environment, such as walk and independent communication, or is the computer and tablet.

CEO Zhang Meng of the Cambrian: the capacity for decision-making behavior mode and with a certain amount of human thinking machines. (Why is a quantitative? ) If 100% human property that should be called a robot. OtterBox iPad Mini Case

Speed technology CEO Chen Zhen: robot, from the industry’s perspective, I think the most broad definition of a robot can replace a person engaged in the work of the machine. We are now deriving from both industrial robots, service robots, educational robot, one of the biggest features is the ability to produce life, help or replace human labor, related functions.

Lei feng’s network: robot program?

Cloud motion technology CEO Tao: hardware body mind core: algorithms, models, and data. Each stage will be different. Before in the absence of large amounts of data, algorithm is important. With the increase of data, will continue to train the model. Robots will become more and more clever.

CEO Misa Rokid robot: program is part of it, including hardware, sensors, identification, understanding and feedback, any intelligent and programs, but it certainly is not the only program.

Read surface technology CEO Zhao Jinglei: robot will became human life in the next platform level of intelligent Terminal, from intelligent terminal for, it contains two chunks, a is intelligent, intelligent can through program to description, and we by said of program more of is refers to artificial intelligence algorithm, including voice, and semantic of understanding and Visual of calculation and understanding, these imitation human cognitive of calculation composition has robot of core, is program this block. In addition, as a Terminal, it will have a wide variety of hardware carrier, overall, it has to have the skills, but it could also have no hardware support, may be run on a PC, such as science fiction movies of the robots in Her “Samantha”.

Lei feng’s network: robot must be someone’s property?

OtterBox iPad Mini Case

CEO Misa Rokid robot: it does not have, such as humans are unable to perceive infrared light. Closer to human perception, of course, easier to understand. Robot I actually advocated not effects too are property of its own, emphasis on feedback interactions of human good.

Speed technology CEO Chen Zhen: telling robots I think should start from the function, he can replace people doing things, rather than in accordance with the principles of people, to define its shape, Visual, tactile. Because such as vision, machine vision and human vision has the same place, but the principle difference God.

Lei feng’s network: that best represent your idea of robot images

CEO Misa:viki Rokid robot (based on the United States’s most famous science fiction and popular science writer Isaac Asimov’s short fiction of the I, robot (I,Robot) super computer in the film adaptation of the same name)

The problem what is the robots, we come to hear entrepreneurs say

Speed technology CEO Chen Zhen: I feel most is inside the movie AI robot kid David.

The problem what is the robots, we come to hear entrepreneurs say

Editorial: can see you for defines the key is automation and robot localization, current robots have been automated, but independence is still in the exploratory stage, once independent, VIKI and David have a chance to get out of screen in our lives, all this and how far from reality.

August 12 to 13th, by the China computer Federation (CCF), Lei feng’s undertaking of the whole network CCF-GAIR global artificial intelligence and robot Summit formally convened its meeting in Shenzhen. The theme of this Summit was “found the future connected innovation” focus on artificial intelligence, robotics, unmanned aerial vehicles, automatic driving in four main areas, under trend of the intelligent community, aims to explore the cutting edge of smart technology and its commercial applications, promote the exchange of the depth of the research. The Summit invitation to the world’s top academic experts, Star team and venture capitalists, including 8 academicians, 25 leading edge innovative enterprises, 100 leaders of technological innovation, and 1200 industry elite. At the Summit, there will be more discussion about the robot, so stay tuned!

Extras Largest dimension of machine learning framework

This article from: company data


Machine learning model to support large dimensions, its data platform with Hong Kong University of science and technology to develop a distributed computing framework for machine learning–Angel 1.0.

FENDI plus case

Angel is a proprietary machine learning system developed using Java language, users can use the Spark, MapReduce, use it to complete a machine learning model training. Angel has supported the SGD, ADMM optimization algorithm, and we also provide some commonly used machine learning model; however, if the user has custom needs, optimization algorithms can also be provided in our top package model with relative ease.

Angel Chukonu as a network solution of the Hong Kong University of science and technology in high dimension parameters of machine learning in the process of updating, targeted to lag computing task parameter speed, on the whole, shorten the time of machine learning algorithms. This innovative use of the Hong Kong University of science and technology Professor Chen and his research team developed the perceived upper-level application (Application-aware) network optimization solutions, as well as the large-scale machine learning research programme led by Professor Yang Qiang.

In addition, Bin Cui, Peking University Professor and his students also participate in the Angel project research and development.

In the actual production, Angel in the tens of millions of levels feature latitude conditions SGD performance Spark several times is a mature open source systems to dozens of times times. Angel has been in Tencent video testimonials, Canton stop precision recommended business practice, and we are currently expanding the application range of Tencent, the goal is to support enterprise-level large-scale machine learning tasks such as Tencent.

The overall structure

Angel in the overall schema reference to Google DistBelief. DistBeilef was originally designed for deep learning, it uses the parameter server to address the huge model update problem in training. Parameter Server also can be used in non-deep learning in machine learning models, such as the SGD, ADMM, LBFGS optimization algorithm on each iteration in the face of billions of update scenario, the need to expand performance parameters distributed caching. Angel supported BSP during operation, SSP, ASP three calculation model, which SSP is validated by the EricXing in the Petuum project at Carnegie Mellon University model in machine learning that particular operations scenario to enhance reduce the convergence time. System has five roles: FENDI plus case

Master: responsible for the application and allocation of resources, and task management.

Task: responsible for task execution, in the form of thread.

Worker: independent process runs on the Yarn in the Container, is a container Task execution.

ParameterServer: built with the start of a task, the task ends and the destruction of, is responsible for the update task parameters in the training process and storage.

Extras! Largest dimension of machine learning framework

WorkerGroup is a virtual concept, made up of several Worker, the metadata maintained by the Master. Considering for the parallel development of the model, all running Worker training in a WorkerGroup data is the same. Although we provide some general models, but does not guarantee meeting demand, and user-defined model can achieve the common interface, formally equivalent to MapReduce or Spark.

1) user friendly

    1. automated data segmentation: Angel system provides users with a capability to automatically cut the training data, user-friendly data parallel computing: system default Hadoop FS interface is compatible, the original stored training samples in support of the Hadoop distributed file system FS interface such as HDFS, and Tachyon.

    2. rich data management: sample data is stored in a distributed file system, system to be read from the file system into calculation calculation process is cached in memory to speed up iterations; if the in-memory cache of data is temporarily saved to the local disk without communications to the distributed file system again.

    3. the rich library of linear algebra and optimization algorithms: Angel also provided an efficient vector and matrix computation library (sparse/dense), and convenience to users of free choice of data, parameters, forms of expression. In terms of optimization algorithms, Angel has realized the SGD, ADMM; models support the Latent DirichletAllocation (LDA), MatrixFactorization (MF), LogisticRegression (LR), Support Vector Machine (SVM).

    4. optional model: we mentioned in the review, Angel of the BSP,SSP,ASP model parameter server can support.

    5. more granular fault-tolerant: fault-tolerance in the system is divided into Master of fault tolerance parameter server disaster recovery, snapshot cache within the Worker process parameters, fault-tolerant RPC calls.

    6. friendly task to run and monitor: Angel also has friendly run support tasks run patterns based on Yarn. Meanwhile, Angel’s progress also facilitates the user to view the Web App page clusters.

2) parameter to the server

In actual of production environment in the, can intuitive of feel to Spark of Driver single points update parameter and broadcast of bottleneck, although can through linear expand to reduced calculation Shi of took, but brings has convergence sex declined of problem, while more serious of is in data parallel of operation process in the, due to each Executor are keep a full of parameter snapshot, linear expand brings has n x parameter snapshot of flow, and this flow concentrated to has Driver a a node Shang!

Extras! Largest dimension of machine learning framework


Seen from the diagram, task in machine learning and Spark even more machine resources are not used, the machine only under certain smaller scale can bring out the best performance, but the best performance is also not ideal.

Using parameter-server scenario, Spark and we made the following comparison: there are 50 million on a data set of training samples using SGD solution of logistic regression models, with 10 nodes (Worker), for different dimensions of each iteration of each characteristic and comparison of overall convergence time (Angel using the BSP model here).

Extras! Largest dimension of machine learning framework

Visible through the data, model Angel contrast Spark advantages are more apparent.

3) memory optimization

During operation to reduce memory consumption and improve operational convergence within the single process uses asynchronous lock-free Hogwild! Pattern. N Task with in the course of an operation if the parameters in the operation remain a separate snapshot memory overhead for parameters n times, model dimensions are consumed when the more obvious! Optimization algorithm of SGD, the actual scene, the vast majority of cases are sparse training data, so update conflicts greatly reduces the probability of, even if the conflict in the gradient is also not entirely to the poor direction, after all, are moving in the direction of the gradient descent updates. We use the Hogwild! mode, so that more than one Task within a process share the same parameter to the snapshot, reduce memory consumption and improve the rate of convergence.

4) network optimization

We have two main point of optimization:

1) process Task parameters after the operation smooth merger pushed the update to update the parameters of the server, which reduces consumption upstream of the machine where the Task, have cut down consumption parameters the server while reducing the push update bottleneck during peak times;

2) further network optimized for SSP: because SSP is a semi-synchronous operation coordination mechanism, in a limited window to run a train, when it reached the edge of the node, node tasks have to be stopped to wait for the slow update parameters. To solve this problem, we flow through the network to speed up the slow process of redistribution of the nodes. We give the slower node with higher bandwidth; accordingly, fast working nodes less bandwidth. In this way, fastest node number of iterations and the slowest node to control the gap, reducing the window be broken (waiting for) the probability, which reduced the working nodes idle wait time due to SSP window.

As shown in the following figure, in 100 million 30 rounds result dimensions, iteration test, you can see the Chukonu of idle waiting for the cumulative time is greatly reduced, was 3.79 times times.

Extras! Largest dimension of machine learning framework

Figure below shows before and after optimization of execution time, with 50 million model dimensions, for example, 20 server nodes and 10 parameters, Staleness=5, the implementation of 30 rounds of iteration. As can be seen, open Chukonu average after completion of the round in 7.97 seconds, compared to the original tasks than average 15% increase was 9.2 seconds per round. FENDI plus case

Extras! Largest dimension of machine learning framework

In addition, the accelerated target node can slow the slow nodes are more likely to get the most recent parameter calculation model for comparing the original SSP, convergence has improved. As shown in the following figure, also is targeting 50 million dimension model under the SSP results evaluation, native Angel task after 30 rounds of iterative (276) loss reached 0.0697, opened after Chukonu, in the 19th round of the iteration (145) have reached a lower loss. From this particular scene is a close 90% of convergence speed boost.

Extras! Largest dimension of machine learning framework

Follow-up plan

Next, the team will expand the size of the application, at the same time, the project team has continued to develop the next version of Angel, next version will be further optimized in parallel model. In addition, the project team is planning to open source Angel, we will follow the right time to open.

This data by Tencent exclusive reprinted, for reprint, please contact the original author.