[AI bulletin 20220211] the hard core up owner has built a lidar and detailed AI accelerator

The embedded AI

Detailed explanation AI Accelerator （ One ）：2012 Year of AlexNet What did you do right ？

original text ：

https://mp.weixin.qq.com/s?__biz=MzA3MzI4MjgzMw==&mid=2650835829&idx=3&sn=b217aeade8dde1d08159bcf58eed6e57&chksm=84e54f0bb392c61dae9401257fad2821a2edaa7345bd5a565e3c1f82f854c7362c6d8fb53e41&scene=21#wechat_redirect

AI、 machine learning 、 The concept of deep learning can be traced back to decades ago , However , They have really become popular in the past ten years , Why is that ？AlexNet The basic structure of CNN There is no essential difference in architecture , Why can it be a blockbuster ？ In this series of articles , Former apple 、 philips 、Mellanox（ Now it belongs to NVIDIA ） The engineer 、 Dr. Princeton University Adi Fuchs Try from AI From the perspective of accelerator, we can find the answers to these questions .

Detailed explanation AI Accelerator （ Two ）： Why is it AI The golden age of accelerators ？

original text ：

https://www.jiqizhixin.com/articles/2022-02-10-6

In the last article , Former Apple Engineer 、 Dr. Princeton University Adi Fuchs Explained AI The motivation for the birth of accelerators . In this article , We will follow the author's ideas to review the whole development process of processors , have a look AI Why can accelerators become the focus of the industry .

This is the second in a series of blogs , We come to the key of the whole series . When promoting a new company or project , Venture capitalists or executives often ask a basic question ：「 Why now ?」

To answer this question , We need to briefly review the development history of processors , Look at what major changes have taken place in this field in recent years .

Real full stack Engineers ！B Station hard core UP The Lord built a lidar himself

original text :

https://www.jiqizhixin.com/articles/2022-02-08-16

Laser radar （LiDAR） It is the abbreviation of laser detection and ranging system , At present, it is widely used in the fields of driverless and sweeping robots . On the one hand, this wide application benefits from the performance improvement of lidar , On the one hand, it also benefits from the decline of its cost .

According to the scanning method , Lidar can be divided into MEMS type 、Flash type 、 Phased array 、 Mechanical rotary ; According to the number of lines , It can be divided into single line type and multi line type ：

Come back , Is it possible to make a self-made laser radar ？B A hard core of the station UP Lord 「 Technology house that doesn't want to live 」 I really achieved this small goal .UP I spent more than half a year in my spare time , With a laser ranging sensor, a single wire mechanical rotary laser radar is assembled , The total includes hardware design 、 The structure design 、FPGA Development and 3D Print a few steps . about UP For the Lord , This is not a new challenge , It can only be regarded as 「 I reviewed what I learned before 」：

Machine learning gains quantum acceleration

original text : 

https://www.jiqizhixin.com/articles/2022-02-11-6

Machine learning gains quantum acceleration

In order to make Valeria Saggio（ Quantum physicist at MIT ） Start the computer in her former Vienna Laboratory , She needs a special crystal ; The crystal is about the size of her fingernails .Saggio Will gently put it into a small copper box , A miniature electric oven , Heat the crystal to 77 Fahrenheit . Then she turns on the laser , Bombard the crystal with a beam of photons .

This crystal , At this precise temperature , Will split some of these photons into two photons . One of them will go directly into a light detector , Its journey is over ; The other will enter a micro silicon chip —— A quantum computing processor . Micro instruments on the chip can drive photons along different paths , But in the end, there are only two results ： The right way and many wrong ways . Based on the results , Her processor can choose another path and try again .

This sequence feels better than Windows More like Rube Goldberg, But the goal is to let quantum computers learn a task by themselves ： Find the right way . about Saggio Come on , This project is similar to trapping robots in a maze . The computer must learn the right path , Without knowing in advance where to turn . It's actually not difficult —— An ordinary classic computer , You can force yourself through dead ends and lucky guesses . but Saggio Want to know ,「 Can quantum mechanics help ？」 last year , Her team proved this .

AI hotspot

Limit overtaking on curves 、 Defeat human top players , SONY AI The racing driver boarded Nature cover

original text ：

https://www.jiqizhixin.com/articles/2022-02-10-4

stay 《GT The car 》 Beat several of the world's top E-sports racers , SONY AI Developed a super powerful car AI agent .

From chess to go to poker ,AI Agents outperform humans in many games . Now? , These agents can be used in 《GT The car 》（Gran Turismo） Refresh the highest score .

《GT The car 》 from SCEJ A racing game led by its famous producer Yamauchi Yidian . From 1997 year , This game is made by POLYPHONY DIGITAL Developed racing game . No matter from the game screen 、 Operate the track while driving 、 Number of cars 、 Realism , The system should be as perfect as possible . This game contains more than 50 Tracks , exceed 1000 models , It can be called the Automobile Museum .

Today Sony announced , Its researchers have developed a product called 「 GT Sophy」 Of AI The driver , It can be found in GT Beat the top E-sports racers in human race for several consecutive rounds . Relevant papers are published Nature cover .

The model is ten times larger , Performance is improved several times ？ Google researchers have done some research

original text : 

https://mp.weixin.qq.com/s/qZ_k9SX0pP7YzJL5tyYrmA

Make a big model , As big as a training to spend millions of dollars to buy computing power , The kind that has no money to retrain , Will you spend money unjustly ？

With the deep learning, the volume of the model is getting larger , Any form of super parameter adjustment will become very expensive , Because each training run can cost millions of dollars . Therefore, some studies aim to explore 「 As the model size increases , Performance improvement 」 Laws . This kind of regular prediction helps to expand smaller scale research to larger and more expensive , But higher performance environments .

As an example of studying the expansion effect , Suppose our goal is to have 3 The width of the hidden layer is exaggerated MLP Medium training ImageNet. We need to 64、128 and 256 Hide size start , And use these to select super parameters , In this case Adam eureka 3e-4 Learning rate of . We also fixed the training length as 30k Weight update , Every time batch Yes 128 Zhang image .

With different hidden layer volumes 8 Performance of different models （ In blue ）. Fitted linear regression （ Black dotted line ） Ideally, it should be possible to predict the loss of a given hidden layer size .

Welfare at the end of article

The computer basics that take you off ！

original text : 

https://mp.weixin.qq.com/s/HrClOEJQWG1bIknctwPLjw

The first is the hard core basics that programmers must know , This is a very introductory classic PDF, After reading, you can have a basic understanding and introduction to computers , It's about training you kernel The basis of , Let's take a look at the outline of the table of contents ：

It basically covers all the basic knowledge of computer , from CPU To the memory 、 Explain what binary is 、 disk 、 Compression algorithm 、 operating system 、 Compilation and other knowledge .

Let's take a look at what the content looks like

  Click to read the original to enter the official website