Shulou(Shulou.com)06/03 Report--

Https://www.toutiao.com/a6695518937334940174/

Text / Editor | there are three words

Some of my classmates always wanted me to write about the interview, but I didn't write it. We will not open the relevant section, because there is no standard, it is easy to cause controversy, and may aggravate everyone's impetuosity and anxiety.

However, there are still some things to say about the interview. The following is divided into two parts.

1 who does the boss like

If I were a boss or an interviewer who is sincere in recruiting excellent people for the company, I would like people with the following characteristics. What we are talking about here is to recruit an algorithmic engineer, a person who can create revenue for the company, and for the sake of simplicity, there is no branch recruitment.

1. Loyalty is very important. A disloyal employee, no matter how capable, can not be asked for. So if you see a resume full of job-hopping and internship resumes, changing internships every two months and changing jobs in half a year, I may not pass this resume unless you give a very objective reason.

2. Coding ability, which is the premise of everything. As mentioned earlier, what we are looking for here is an algorithm engineer in the field of computer vision (CV). The company will not recruit a person who is not familiar with Matlab or python to do project development. I can't do pure algorithm research. I haven't seen a great person who can't realize his idea. The awesome people I have seen are not only good at algorithms, but also have strong programming skills.

When it comes to programming capabilities, there are a few basic requirements:

(1) linux should be cooked.

This is not the era when there were not many open source projects a few years ago, and everyone is still using Matlab and VS simulation under Windows. It has been 9102 years, and Linux is not familiar with it. How to reassure the boss about efficiency.

(2) python is familiar, C++ is good.

If you are not familiar with such a simple language as python, you really have nothing to say. This is a basic course, science and engineering schools have not opened should be very rare, CV in the field of industrial deployment and algorithm optimization can not be separated from C++.

(3) good programming habits.

Although it is said that the code is poorly written, no colleagues will really surprise you with gun, but good programming habits not only improve efficiency, but also look comfortable and will not be despised by others.

The habits mentioned here include: writing more classes and function encapsulation, organizing the project directory structure, naming well, and so on. Can not write how awesome cool code, but to ensure that the code has a good scalability, easy for others to read the migration, the specific requirements will be discussed later.

3. The basis of the algorithm, which determines the potential. No tests? It's okay. Haven't you done any segmentation? Nothing. Learn while doing and follow up quickly. This is the norm in the development of the company.

People's energy is limited, and they haven't done much. But if some of the basics of CNN are not clear, and you don't know the basic concepts of images, that's a problem. Because if you give him a project, you may make some low-level mistakes without knowing it, and the boss will panic and accompany him to check and popularize basic knowledge.

These are the three basic requirements, which must be required by every company. I think if they pass these three tests, they will at least be a candidate who can be trained, and I will be willing to give him the opportunity to enter the next round of competition.

Solid (good foundation), reliable (stable), able to work (able to write code), these three basic premises are more important than anything else, the rest can be learned in the project.

2 what knowledge should be prepared

Next, it's time to talk about what to do. This is why I opened the column "the Road of AI practice". If you are patient, you might as well read on. We also review and summarize the previous articles. There are good reasons for the opening of each article.

1. Why use Linux

In the past, you might have thought that Linux was not a rigid requirement, that you used your own Windows computer, and that you didn't need to share operating systems, hardware and disks with others. But if your team uses the server together, it's impossible not to use Linux.

So this is for children who have not really done daily development work on Linux, to officially enter the AI industry, Linux is the necessary and only operating system, "soft" weapon, I have not heard of any company training models on Windows or Mac.

The deep learning of "AI White body" begins with abandoning the use of windows.

2. Use shell,vim,git to improve development efficiency

Programming habits, work efficiency is very important, very important! A skilled engineer under Linux will be much more efficient than Windows. To improve the efficiency of writing code, we can start from terminal multitasking management, proficient in the use of shell commands, proficient in the use of vim and other development environment, proficient in the use of git commands and other places.

The shell command is the operational basis of Linux, and it is also the beginning of learning to use Linux, and getting familiar with advanced shell commands will bring great efficiency improvement in future work.

Vim is the most commonly used editor under Linux, from a small white to a master can be used, and its column editing, find and replace, automatic completion and other functions are the guarantee of efficiency, perhaps students who switch from the environment such as visual studio will not adapt at first, but as time goes by, they will more and more understand the good of VIM.

Git is a necessary quality for programmers. Learn to maintain a few of your own code bases slowly, and you'll understand when something goes wrong in the future.

"AI White body" Linux works with three axes, shell, vim and git

3. Python fundamentals and programming habits

In the field of machine learning, python can be said to be a clean ride. To learn python, you need to master the basic syntax, including functions, class design, and master a large number of open source matrix libraries such as Numpy.

Is python easy? simple. Is it really easy? Take a look at the projects written by the gods.

"AI White body" the necessary python basis for learning AI

4. Image basis

One of the biggest problems with deep learning is that it is too easy to use. Lead to what image basis and traditional algorithms do not need to understand, can also be done by the project.

But if there is no good image foundation, one day there will be problems that CNN can't solve, or problems that can't be solved alone, you won't know what to do. Many new technologies are inspired by traditional algorithms or combined with each other, do not understand the basis of the image, it is like burying a fixed-time bomb, generally nothing, blow up.

The essential Image Foundation for Deep Learning of "AI White body"

5. OpenCV Foundation

If there is any library in the field of image processing that can not be bypassed, it must be OpenCV, an open source computer vision library that can be called the best computer vision library, not only free to use academically and commercially, but also cross-platform and high performance. The basic contents that need to be mastered include: how to deploy, the familiarity and use of basic data structures, and the understanding of basic modules.

"AI White body" is a necessary OpenCV entry basis for computer vision.

6. CMake compilation

Python is a scripting language, but at present, a large number of AI algorithms are deployed on the mobile embedded platform, which need to use c/c++/ java language. Linux and Makefile are the tools for compiling C code under CMake.

In fact, some python,matlab open source projects also need to be precompiled, and more of them will be understood when they are in the workplace.

"AI White body" can only use Python? Let's get to know about glossy girls and Makefile.

7. Reptile basis

What is the most important thing in deep learning project development? data, of course! You often don't have enough data for actual projects, so you need to figure out how to get it on your own.

The Internet is a great treasure house of all kinds of resources. Learn to use crawlers well, and you will probably become the most "resourceful" person in the era, which is also likely to be the beginning of the success of the project.

The last actual project in this article needs to be used.

"AI White body" study in depth the basis of reptiles that you have to know

8. Data visualization

After crawling the data, it should be processed. A very common method is data visualization. In deep learning projects, the data visualization operations that are often needed include the visualization of original image data, the visualization of loss and accuracy, and so on.

In addition to data visualization, we also need to visualize the model to facilitate debugging and perception.

Data Visualization in Deep Learning of "AI White body"

9. Mathematical basis

Although for most people, you don't need a strong mathematical foundation to do a project, you will need to understand other people's papers, and you will often need to make simple deductions and algorithm improvements.

From linear algebra, probability theory and statistics to calculus and optimization, all need to be mastered. However, the learning of mathematics is a very long process, do not rush to success, and can not be completely learned by following videos or books, the important thing is to use it.

What is the mathematical basis for "AI White body" to enter AI: left-handed matrix theory, right-handed calculus

10. Research direction of computer vision

Study and do the project need a direction, in front of these foundations have been mastered, it is necessary to have a good understanding of the major research directions and characteristics of computer vision, to facilitate their own topics and project plans finalized.

From image classification, segmentation, target detection, tracking, to image filtering and denoising, enhancement, stylization, 3D reconstruction, image retrieval, GANs, I believe there is always something you like or the project involves.

The article "AI White body" reviews the research direction of computer vision.

11. Application direction

The ultimate purpose of learning is to solve practical problems. AI has infiltrated into all aspects of our lives. From self-driving cars, image beauty, to chat robots, financial payments, etc., so it is right to take a good look at the current application of AI in all major fields, this time is not limited to computer vision.

Https://dwz.cn/Gj18HPHd

12. Get to know academic celebrities

If you want to really integrate into the industry circle and keep up with technological development, you must always understand the status of the bosses, who are the weather vane of the development of the industry.

Whether it's academia or industry, whether it's an old master or a young talent, let's think of what we can do together.

We have opened up the project of the boss's research direction, welcome follow.

Https://github.com/longpeng2008/Awesome_DNN_Researchers

Who on earth is the king of AI, the top 10 computer scientists in the world?

13. A brief history of artificial intelligence

After the foundation is laid, the next step is to formally learn the knowledge related to AI, no matter which class or textbook you are in, it is to let you know the sages first.

From the enlightenment of artificial intelligence caused by Turing and machine intelligence, von Neumann and brain-like computing, to the twists and turns of the three waves and the growth history of technology, it is worth reading for everyone in the industry.

"the initial understanding of AI" starts from three ups and downs of artificial intelligence.

14. Neural network foundation

The method of deep learning to study problems is to imitate the brain, which is based on neural networks. From receptive field, to MP model, to perceptron, to back propagation, it is necessary to be familiar with the disadvantages of fully connected neural networks, the characteristics of convolution neural networks, core technologies and advantages, which is the most important basis for deep learning.

Understanding the Neural Network from the beginning of AI-the Watershed between the expert and the layman

15. Breakthrough in the field of understanding

Since learning deep learning, it is necessary to understand the important progress of deep learning.

In the pre-deep learning era, the discovery of visual mechanism, the first convolution neural network Neocognitron, and the popularity of back propagation algorithms promoted the birth of LeNet5 and MNIST data sets.

With the maturity of the new theory, the accumulation of big data and the popularity of GPU, the technology represented by convolution neural network has made a major breakthrough in the basic fields such as image classification and target detection. With the success of AlphaGo, it has planted the seeds of deep learning / artificial intelligence technology in the minds of both industry and laymen, and has been full of vitality ever since.

The important Progress Node of Deep Learning in the Field of Image in the past 20 years of "AI initial knowledge"

16. Activation function

The mechanism of deep learning is imitated in the human brain, and the cells of the human brain need a certain threshold to receive stimulation to produce activity, which is the fundamental origin of the activation function.

The activation function shoulders the improvement of the nonlinear expression ability of the network, from the early smooth sigmoid and tanh activation function, to the later ReLU and various variants of ReLU (LReLU,PReLU,RReLU,ELU,SELU,GELU, etc.), Maxout, researchers have been trying to make the network have better expression ability.

With the development of technology, the use of reinforcement learning and other algorithms to learn new activation functions from the function pool, such as swish, has become the mainstream of current research, and activation functions have also embarked on the road of data-driven.

The activation mechanism seems simple, but in fact it is not easy. We must follow up and understand more.

Activation function of "AI beginnings": from manual design to automatic search

17. Parameter initialization

Parameter initialization, which seems to be a very simple problem, has really trapped the optimization of neural networks for a long time. In 2006, Hinton and others published a paper "Reducing the dimensionality of data with neural networks" in the journal science, which opened the prelude to a new training of deep neural network algorithms, which is still regarded as the era of the third upsurge of artificial intelligence.

From all-zero initialization and random initialization, to standard initialization, Xavier initialization, He initialization, and now thousands of layers of network training has become a reality, initialization no longer seems to be such an important topic, but who says there is no room to think.

What is the beginning of the success of in-depth learning in AI? Parameter initialization

18. Normalization

We always hope that the statistical problems we study can satisfy a fixed distribution, and this will indeed reduce the difficulty of the problem.

In deep learning, because the number of layers of the network is very large, if the data distribution in a certain layer begins to shift obviously, this problem will intensify with the deepening of the network, which will increase the difficulty of model optimization.

Normalization is committed to solving this problem, from data to weight, from one characteristic channel limited in the same sample to all channels of different samples, all kinds of normalization methods in a simple way, elegantly solve the problem that deep learning model training is easy to fall into local solution, and improve the training speed and generalization ability by the way, which is a theoretical and engineering skill that must be mastered.

How much do you know about Normalization in the deep learning model of "AI first acquaintance"?

19. Pooling

The brain learns knowledge by abstraction, and abstracting knowledge from images is a process of filtering and extracting information "from big to small". Pooling, which comes from the visual mechanism, is the process of abstracting information.

Pooling increases the invariance of the network to translation and improves the generalization ability of the network. We are used to using mean pooling mean pooling and maximum pooling (max pooling), although convolution with step size can be used instead.

Although the role of pooling has been questioned by researchers, the pooling mechanism is still a necessary structure in the network, so you must be familiar with it, and the data-driven pooling mechanism is worth studying.

What on earth is the pooling of "AI at first sight" rejected by Hinton,DeepMind and Stanford?

20. Optimization

The learning of the model can only be realized through the optimization method. The optimization of deep learning model is a non-convex optimization problem, although the first-order and second-order methods can be used to solve it, the current random gradient descent SGD and its variants are still the first choice.

Starting from SGD, some focus on improving its optimization speed, such as Momentum momentum method and Nesterov accelerated gradient method, some focus on making different parameters have different learning rates, such as Adagrad,Adadelta and Rmsprop methods, some hope that people can extricate themselves from tuning parameters, such as Adam method and its variants, and some strive to make the convergence process more stable, such as Adafactor method and Adabound method.

No method is perfect, you have to try it when you train.

"AI at first sight" in order to encircle and suppress SGD, there are more than a dozen tricks that everyone has thought about over the years.

21. Generalization ability

It is useless if a model can only work on the training set.

Therefore, we always hope that the model performs well not only for known data (training set), but also for unknown data (test set), that is, it has good generalization ability, which is realized by adding regular items.

From parametric regularization methods that directly provide regularization constraints, such as L1/L2 regularization, engineering techniques such as early termination of training and model integration, and implicit regularization methods such as data enhancement, researchers have invested a lot of energy in this field, which we must pay attention to all the time.

How to increase the generalization ability of Deep Learning Model in "AI initial understanding"

22. Model evaluation

There is no evidence for words, and using data to speak is the correct posture for researchers to compete. How many tasks of computer vision, from classification, regression, quality assessment to the generation of models, of course, we need to master scientific evaluation methods.

Evaluation of the deep learning model of "AI initial knowledge", from image classification to model generation

23. Loss target

The learning of the model needs guidance, which is the responsibility of the loss function, which often has a great impact on the final performance of the model.

This article focuses on summarizing classification problems, regression problems, and generating loss targets used in countering networks, so as to lay a theoretical foundation for everyone to design better optimization goals.

What are the loss functions commonly used in the in-depth learning of "AI beginners"?

24. How to start training your model

Sharpening knives doesn't make mistakes in chopping firewood, and when we start to train our models, we always have to think clearly about something.

The first step is to know what task you are going to do, find competitors to make good expectations, and figure out what kind of data you need. The second step is to determine the framework, benchmark model, and prepare the data. Then the third step is to start the training, from input and output, data preprocessing to maintaining the correct training posture.

This is the experience I have summed up, and it must always be useful.

Ten suggestions of "AI beginners" to do projects for beginners in deep learning

25. Work on a project seriously

In the end, I believe everyone has the basic skills, so let's start a real project to achieve some of the functions of App.

I will not teach you how to finish it here, because this project is not simple at all, it is a very difficult classification project, it cannot be solved by an existing data set and a model, at least the following problems need to be considered, each of which has a lot of engineering details to be solved.

If you can independently complete the entire process of this task, identify 20 kinds of flowers with long and close range, and achieve 95% accuracy, then I believe you are fully capable of applying for a computer vision algorithm engineer.

If you have any problems with the framework, you can refer to our 12 mainstream open source frameworks using git projects in various tasks.

Https://github.com/longpeng2008/yousan.ai

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.