website development austin


The main points of this article

  • We see more and more companies using deep learning algorithms。therefore,We transfer deep learning from innovators to the category of early adopters。What is related to this is,Deep learning is also facing new challenges,For example, deploying algorithms on edge devices and training of very large models。
  • Although the speed used is relatively slow,But now there are more commercial robot platforms。We have seen some applications outside the academic world,But I believe there will be more unsightly use scenarios in the future。
  • GPU Programming is still a very promising technology,But it has not been fully used yet。In addition to deep learning,We believe there are more interesting applications。
  • Use Kubernetes Such technology,Deploying machine learning on typical computer stacks is becoming easier and easier。We see that the continuous tools are being automated to automate the increasing part,For example, data collection and re -training steps。
  • AutoML It is a very promising technology,It can help data scientists focus on the actual problem domain,Instead of focusing on how to optimize the supers。
website development austin

InfoQ Edits discuss each year AI、ML And the current status of data engineering,So as to identify as a software engineer、The key trend that architects or data scientists should pay attention to。We organize our discussions into technology to use curves and add related comments,To help readers understand the evolution of things。We also explored part of the development of the roadmap and skills,What should you consider。

We also recorded these discussions for the first time intoInfoQ PodcastSpecial program on it。Kimberly McGuireYes Bitcraze Robot engineer,Daily work is dealing with autonomous drones,He joined the editorial department to share his experience and opinions。

Deep learning transferred to the early adopted

(website development austin)Although deep learning is in 2016 Only the year began to cause our interest,But now we decide to make it from the innovator(Innovator)Category transfer to early adopted(Early Adopter)。We see that there are two main frameworks in deep learning,SeparateTensorFlowandPytorch。Both are widely used in the entire industry。We should admit,PyTorch He is the leader in the field of academic research,and TensorFlow Business/Leader in the corporate field。These two frameworks have maintained a considerable balance in terms of function,So the specific framework depends on your requirements in terms of production performance。

We have noticed,More and more developers and organizations are collecting and stored their data,Follow this way,That is easy to be processed by deep learning algorithms,In order to“study”Something related to business goals。Many people set up their machine learning projects for deep learning。TensorFlow and PyTorch Establish an abstract layer for various types of data,And incorporate a large number of public datasets into their software。

(website development austin)We still see,The scale of the data set used for deep learning is increasing significantly。We saw,The next challenge is distributed training achieved with distributed data and parallel training。There are examples of this framework FairScale、DeepSpeedandHorovod。That's why we will“Large -scale distributed deep learning”The reasons for introducing the theme list of innovative categories。

Another challenge we see in the industry is related to the training data itself。Some companies do not have large data sets,This means that they can benefit a lot from pre -training models using their specific fields。Since creating a data set may be a high -cost work,Choosing the right data for the model is a new challenge,The engineering team must learn how to solve this。

The edge deployment of deep learning applications is a challenge

(website development austin)Currently,Run on the edge device AI Still challenges,Such as a mobile phone、Raspberry Pi,Even smaller microprocessors。The challenge here is to deploy the model trained on the large cluster to a small hardware。To achieve this, the technology to be dependent is the quantification of the network weight(Use less bits for network weights)、Network trimming(Remove the weight of little contribution)And network refining(Train a smaller neural network to predict the same content)。E.g,This can be through Google TensorFlow light and NVIDIA of TensorRT to fulfill。When we narrow the model,Sometimes I really see the decline in performance,But how much performance decreases and is this a question,It depends on the application。

Interestingly,We see that some companies are adjusting their hardware to better support the neural network。In Apple devices and the core of tensor(tensor core)of NVIDIA In the graphics card,We all saw this。Google New Pixel There is also a number of chips for mobile phones,You can run a neural network locally。We think this is a positive trend,It will enable machine learning to use more environment than now。

The business robot platform for limited applications becomes more and more popular

In the family,Robot vacuum cleaners are already very common。A new robot platform is becoming more and more popular,It isSpot:Boston Dynamics Walking robot。It is being used by the police station and the army for daily monitoring such scenes。Although such robot platforms are successful,But they can only be used within a limited range,And in a very limited scene。However,With the improvement of artificial intelligence capabilities,We hope to see more use cases in the future。

A robot that is moving towards success is autonomous driving car。Waymo And other companies are testing cars without safety drivers inside,This means that these companies are full of confidence in the capabilities of these vehicles。We believe,The challenge facing large -scale deployment is to expand the feasible area of these vehicles,And prove that these cars are safe before going on the road。

GPU and CUDA Programming allows the problem to perform parallelization

(website development austin)GPU The programming method allows the program to perform large -scale parallel tasks。If the programmer's goal can be achieved by dividing a task into many unreasonable kid tasks,Then this program is suitable GPU Programming。Unfortunately,use NVIDIA the company's GPU Programming languageCUDAProgramming,It is still difficult for many developers。There are some frameworks to help us,Such asPyTorch、NumbaandPyCUDA,They should make this programming method easier to enter the general market。Now,Most developers are in use GPU Implement deep learning application,But we hope to see more applications in the future。

Semi -supervised natural language processing performs well in the benchmark test

GPT-3 And other similar language models“GM natural language API”The performance is very prominent。They can handle a variety of inputs,And it is breaking many existing benchmarks。We saw,Semi -supervised(semi-supervised)The more the data is used,The final result is better。They not only perform well on normal benchmarks,And at the same time, summary of many benchmarks。

The architecture of these neural networks,We see people from LSTM Such recursive neural network turns transformer Architecture。The training model is very huge,Use a lot of data,And spend a lot of money for training。For the funds and energy consumed by these models,Treating some related criticisms。Another problem of a big model is the reasoning speed of reasoning。When real -time application is realized for these algorithms,They may not be fast enough。

MLOps and Data ops It can be easier to achieve training and re -training algorithms

(website development austin)We saw,All the main cloud suppliers support the universal container arrangement framework,Such asKubernetes,They are increasingly integrated ML Good support for usage scenarios。This means that we can easily deploy the database as a container on the cloud platform,And expand and expand it。One of the benefits of doing this is,It has built -in monitoring。One of the tools worth noting isKubeFlow,It can be in Kubernetes Coordinate and complex workflow。

About deployment algorithms on the edge,We saw the improvement on tools。for exampleK3s,This is suitable for edge Kubernetes,besidesKubeEdge,It and K3s Different。Although these two products are still in the initial stage,But they are expected to improve the deployment of container -based artificial intelligence on the edge。

We also see some of the complete support ML Ops Products of life cycle are appearing。One of these tools isAWS Sage maker,It can help us train the model easily。we believe,finally ML It will be integrated to complete DevOps Life cycle。This will create a feedback loop,We deploy an application,Monitoring application,And look back before re -deployment to make some changes before re -deployment。

AutoML Allow ML Part of the automation of life cycle

We see the so -called“AutoML”The people increased slightly:In this technology,Part of the life cycle of machine learning will be automated。Programmers can focus on obtaining the correct data and general concepts of models,The computer can find the best super -reuse(hyperparameter)。Now,This is mainly used to find the architecture of neural networks,And finding the best super -reuse to train the model。

We think this is a good progress,Because it means,In terms of transforming business logic into a format that can be solved by machine learning,Machine learning engineers and data scientists will play a greater role。We think this effort makes it more important to follow the experiments that we are doing。pictureMLflowSuch technology can help track experiments。

all in all,We think the problem space is from“Find the best model to capture your data”Turn“Find the best data to train your model”。Your data must be high -quality,Your dataset must be balanced,And it must contain all possible marginal scenes of the application。To do this, it is mainly implemented by hand.,And you need to have a good understanding of the problem area。

What do you need to learn to become a machine learning engineer

We believe,In the past few years,Machine learning has also changed in education。Starting from the classic literature may no longer be the best way,Because there have been too many progress in the past few years。We recommend choosing a deep learning framework to get started,Such as TensorFlow or PyTorch。

It is a good idea to choose a dedicated discipline。exist InfoQ,We divide the discipline into the following categories:Data scientist、Data engineer、Data analyst or data operation and maintenance。According to the major you chose,You have to learn more about programming、Statistics or neural networks and other algorithm knowledge。

Be InfoQ Editor,What we want to share is,Suggestion to participateKaggleContest。You can choose a question in the field you want to know,Such as image recognition or semantic segmentation。By creating a good algorithm and in Kaggle Submit the result,You will see your solution and others participating in the same competition Kaggle What level of users is at what level。In this way, you will have motivation Kaggle Get a higher ranking on the ranking,Usually the winners of the game will write down their winning methods after the game.。so,You will continue to learn more skills,Therefore, it can be directly applied to your problem area。

In the end, but the same important thing is,InfoQ There are also many resources。We often publish the latest and most important news about machine learning、article、Speech and Player。You can also see how our article is successfully applied as a machine learning engineer。At last,Please participate 11 MonthlyQCon plusMeeting,Participate“ML everywhere”Theme of。

About the Author:

(website development austin)Roland Meertens Is a computer visual engineer,exist Autonomous Intelligent Driving Intelligent computer visual algorithms engaged in autonomous vehicles。Before,I have studied natural language processing(NLP)Deep learning method of problem、Social robotics and computer vision of drones、Machine learning and computer visual issues。The interesting thing he does is the translation of neural machine、Capital avoidance of small drones,And social robots serving the elderly。Except InfoQ Published news about machine learning,Sometimes he is also in his blog and twitter( an article。In my spare time,He likes to run in the woods,And participate in obstacles。

Kimberly McGuire Now I am at Bitcraze AB work in company,Poor software developer。2019 year,She obtained a doctorate degree in the School of Aerospace Engineering, Dalph University of Science and Technology, Dutch。The theme is about“Explore bee colony with pocket drone”。McGuire Those who have been calculated in the calculation capacity MAV Complete the biological inspiration of interior exploration,These MAV Can be placed on the palm。besides,She has artificial intelligence to artificial intelligence(embodied artificial intelligence)Extensive interest,And strive to keep up with the latest development。

Srini Penchikala It is a senior in Austin, Texas IT Architect。He is in the software architecture、There are more design and development in design and development 25 Annual experience,Currently focusing on Yun native architecture、Micro -service and service grid、Cloud data pipeline and continuous delivery。Penchikala WriteBig-Data Processing with Apache Spark,And write with others Manning Published“Spring Roo in Action”。He often speaks at the meeting,Is a big data trainer,And published a number of articles on various technical websites。

Raghavan "Rags" Srinivas (@ragss) Is a architect/Developer evangelist,The purpose is to help developers establish a highly scalable and available system。Be Rackspace the company's OpenStack Promotioner and solution architect,He is constantly facing the challenge from low -level infrastructure to high -level application issues。The field he is concerned about is a distributed system,Specializing in cloud computing and big data。exist Hadoop、HBase and NoSQL Early stages,He has been engaged in related work。He has obtained many times JavaOne rock star title。

Anthony Alford Yes Genesys Development team manager,He is engaged in several artificial intelligence and artificial intelligence related ML project。In terms of design and constructing scalable software,He has more than 20 Annual experience。Anthony Possessing a PhD in Electronic Engineering,Professional is intelligent robot software,Interactive interaction with artificial intelligence and artificial intelligence SaaS Various issues in the field of business optimization prediction analysis have studied various issues。

View original English text:

Article: AI, ML and Data Engineering InfoQ Trends Report - August 2021