edit | Mingming

1 month19 day, The innovator conference in Geek ParkIF2018 Scene,Google Brain Chen Zhifeng, chief engineer, publishes:《 The answer starts with defining the question ——TensorFlow
What can be used for?》 Speech, SharedGoogle Brain Research direction in the last year to two years, And inTensorFlow Some work on, Achievements, Progress. The following is the actual record of the speech:

Deep learning

Deep learning is very popular these years, stayGoogle In the search traffic of, Deep learning in the past7-8 Year time, Increased by about100 times, From this aspect, it also reflects the academic and industrial attention to this technology, It's growing rapidly.

Everyone wants to know, What is deep learning? What can it do for us? How to apply this technology to the actual scene?

Deep learning is not a sudden technology field, Its core algorithm is neural network. Neural network is a model of machine learning, The main characteristics of this model are, It can fit any mathematical function, Especially using a similar method of big data iteration to train the model.

With the popularization of big data, The development of computer hardware computing power, And the breakthrough of the algorithm itself, All of a sudden, This deep learning technology with neural network model as the core, In many application fields, it can greatly improve the tasks that software systems need to complete in the past.

You must be familiar with the most classic application fields, Image recognition. since2012 Since the year, In this field, The error rate of automatic image recognition is from20% All the way down to4% About, It's beyond an average person's ability to distinguish images.

Google We started to apply deep learning technology to many products many years ago, In the process, We also iterate, Several generations of software systems supporting deep learning have been developed, Which eventually led us to2015 year10 The month is open source.TensorFlow, Hope to further promote the application and research of deep learning.


TensorFlow Now it has evolved into a complete open platform for deep learning software.

Platform supporting multiple hardware

For instance, It supportsCPU,GPU Training platform of data center built in a mixed way, It also supports models that train data centers, Relatively convenient deployment to different mobile applications, Can support similarGoogle Independently developedTPU processor.

This multi platform support, Can help the most users and application scenarios, We are also very grateful for the support from many industry peers, For example, in the United StatesIntel And NVIDIA are helping us optimizeTensorFlow Performance on their respective hardware.

Support multiple development environments

Platform supporting multiple hardware is the foundation,TensorFlow A long-standing goal, To help as many developers as possible, Able to make use of deep learning technology, Ultimately, the vast number of users can benefit from this capability, Based on this idea,TensorFlow Always attach great importance to the support of various programmer development environments. For instance, Developers can use it in the main development environmentTensorFlow.

TensorFlow The application and promotion in our company is very comprehensive and thorough, For instanceTensorFlow Help earlyGoogle Core business of( Search and advertising), Implemented the model of deep learning, And reflect their performance in the core business.

In spam filtering, We used it, tooTensorFlow Training model, At the same time, on the android app recommendation, It's on the line.TensorFlow Models, etc. Quite a lotTensorFlow Applications are all happening in the background, Most users may not have a direct experience.

Deep Mobile

Let me give you a few examples on the mobile end, Deep learning has directly affected thousands of users.

Android phone selfie

For instance, A selfie feature has been added to the latest version of Android Phones, This selfie function is a visual model trained through deep learning, Very accurate separation of foreground and background pixels, Processing foreground pixels and background pixels respectively, In this way, the background can be virtualized.

To achieve this function, Traditionally, mobile phone manufacturers need to add a second camera, This will increase the cost of mobile phones, At the same time, the existing mobile phones are not easy to achieve such an effect, Through the new algorithm, We can achieve something that might have been very expensive in the past.

Smart speakers

Same as image processing, Speech processing is another area that has been profoundly changed by deep learning, Speech recognition uses neural network earlier than image processing.

Product aspect, Smart speakers have become the main voice players in recent years, One of the big reasons is that deep learning algorithm greatly reduces the technical threshold of speech recognition and speech generation, May have been needed in the past20—50 A task that can only be completed by a team of doctors, Now download an existing model and customize it, You can achieve the same effect. General machine learning framework, Can help more developers, Develop voice applications suitable for specific application scenarios.


Another example, Now you can take a picture with your cell phone, Mobile software can automatically recognize the text in the graphics, Translate words into another language, Seemingly simple application, In fact, it's a natural combination of image technology and machine translation technology.

More than a year ago, I'm lucky to be part of Google, A project to upgrade the past translation system to a neural network-based system, That upgrade greatly reduced the error rate of machine translation. Translation between some of these languages, It can almost achieve the effect of artificial translation

Message auto reply function

Using deep learning technology, We can not only greatly improve the function and performance of existing products, We've also developed new features that were hard to imagine in the past, For example, the function of automatic email reply, Email software on Android can analyze users' email.

For example, you received an email from your friend this morning,“ Do you want to eat somewhere in the evening”, In most cases, there are only three possible answers: I'll be there on time; I'm sorry I can't come without time; I'm free, But I might need to be late.

Now we will automatically give you three choices, In this way, I can see the mailbox on the road, One click to complete email reply, It greatly facilitates the efficiency of handling email on mobile phones.

TensorFlow Applications in other fields

Through many previous application examples, We can see, A general deep learning framework, It can help many existing Internet applications improve their intelligence level, But we can also see,TensorFlow Such a framework for deep learning, It can also be applied in many other fields.


For example, our colleagues in London, The past two yearsAlphaGo Work, That's becauseTensorFlow Help of framework itself, becauseTensorFlow Can take advantage of large computer clusters, Support for the latest accelerators,AlphaGo The team can focus more on Algorithm Research.

Intelligent medical treatment

There are also people in our group who are doing intelligent medical work. BecauseTensorFlow It's a general framework, They can easily reuse existing image recognition models or natural language processing models, Specific application areas and data, Retrain and fine tune the model, You can get more than95% Accuracy rate, That's more accurate than the average ophthalmologist91% Level.


In terms of autonomous driving, We areAlpha Waymo Colleague, Also in useTensorFlow Continuous improvement of depth models in automated driving systems, Including segmentation of traffic scenes, Radar signal processing, etc.

Baby food production

We are particularly pleased, By way ofTensorFlow Open Source, The threshold for us to use deep learning technology has been greatly reduced. In the past two years, we have seen many industries that have nothing to do with the Internet, Also began to try to use deep learning technology and methods, For example, this company is a baby food production company, IntroducedTensorFlow Trained intelligent system, This intelligent system can classify the raw materials of baby food, Get rid of some rotten apples and bananas more accurately, This can accurately control the quality of baby food.

scientific research—— Space

We are also very excited to see,TensorFlow It is also used in scientific research other than computer science, For example, last month Nasa announced a study, NASA has a Kepler program, Their scientists and our colleagues jointly developed aTensorFlow Model.

The goal of the Kepler project itself is to continuously observe the changes in the brightness of stars in space through telescopes, Discovery of planetary systems outside the solar system, Finally, we hope to find another planet suitable for human habitation. At present, the program has accumulated tens of billions of observation data, A few months agoTensorFlow Model, To help scientists find out2500 Kepler beyond light years90 The eighth planet in the galaxy.

scientific research—— ocean

People not only look up at the stars, And we're looking back at our planet, In this application example, For scientists in AustraliaTensorFlow Developed image recognition model, In tens of thousands of marine aerial photos, Quickly and accurately locate large marine mammals in need of protection, For example, cherish animals, manatees.

scientific research—— Jungle

It's also used by scientistsTensorFlow Using voice processing technology to protect birds. They installed a lot of microphones in the jungle, Collect the sounds of birds, The model can accurately estimate the number of birds in a forest, So as to protect them more precisely.

Creating music

Our group and colleagues have made some very interesting applications, They're trying to use deep learning technology to create music, It's interesting, The music created by these music, I've also received professionalDJ Affirmation.

Increase support for Chinese market

since2015 Since open source in, We've been trying to increase thatTensorFlow Input, Recently, we began to increase our support to the Chinese market, Let's say we're building aTensorFlow Chinese website of,TensorFlow And the developers of, meanwhile, We've seen more than1000 Many people are involvedTensorFlow Development, This is a very active community.

in addition, Our data also shows, Up to now, the world180 Many countries, Various users have downloaded more than1000 Ten thousand timesTensorFlow Development Kit, This showsTensorFlow Application development is also very common.

The picture above also tells you, There are a lot of developers in China who are constantly paying attentionTensorFlow, In fact, We are building partnerships with many Chinese companies, Actively support and help them better useTensorFlow.

For example, the interior of JD has been builtTensorFlow Training platform, For developing images, Models related to natural language, And use them in customer service advertising and other fields. Xiaomi is trying a similar technical route, Support their ecosystem for a variety of special applications. Youdao notes of Netease, Netease translator Jun also used itTensorFlow Models of vision and language.

Eager Pattern

Thank you very much to our users and partnersTensorFlow Feedback,TensorFlow Efforts are also being made to develop new functions, Last year we developed aTensorFlow Pattern, This mode can be more conducive to front-end development, Easy to debug, Support more dynamic programming modes.

There's a very short program here, In this program, You should be able to see that the main feature of this pattern is in the front-end process, It can more directly reflect the flow of program algorithm logic itself, We believe that this will be especially helpful for rapid prototype development and debugging.

Lite Pattern

in addition,TensorFlow The main projects promoted in the past year areTensorFlow Lite Pattern, This mode is specifically for mobile动和嵌入式应用场景打造的机器学习平台,它的目标是希望把在云端训练的机器学习模型,更加简单,高效的迁移到移动端上进行部署.





2018 年了,该不该下定决心转型AI呢?





干货 | AI 工程师必读,从实践的角度解析一名合格的AI工程师是怎样炼成的


详解 | 如何用Python实现机器学习算法

经验 | 如何高效学Python?