** Guide language: ** Non-AI professional technical personnel to transition to AI technology, or as a student to learn AI technology development, for everyone with such aspirations and experience, they hope to see the growth experience of AI technical talents, give their own real experience to share.

preface

To follow the thinking of Samuel Johnson, “When a man is tired of learning technology, he must be tired of IT; Because only with continuous learning can you have everything the IT industry brings to you, including money.” This is the reality of the IT industry, no one can live on their laurels for a long time, especially AI technology. Recently, I was reading Londoners, which tells the personal feelings and stories of over 200 Old and new Londoners about the city of London. I felt that I should write an article about the growth of AI talents, so this article came into being.

For those who are not AI professionals transforming to AI technology or learning AI technology as a student, everyone with such aspirations and experiences may hope to see people with similar experiences and share their real experiences.

Today, I found several of my colleagues, including me, a total of five, we are very representative, one by one to introduce:

  • MaiKeZhou: I graduated from Zhejiang University in 2004 with a master’s degree in computer science. I began to learn programming at the age of 15, using Basic language, and mainly wrote C language during my study. When I graduated in 2004, I wrote JSP code (a way to embed Java language in HTML code). Later, I entered the field of big data technology and engaged in AI platform product research and development in the recent 4 years.
  • Mr Qiu: Master degree in algorithm related specialty, working for five years. In 2013, I entered Hangzhou Dianzi University to study artificial intelligence and electronic Information Technology. My master’s degree mainly focuses on embedded and image algorithm. After graduation, I entered Dahua Company in the security industry and then Huawei, where I have been engaged in computer vision, algorithm transplantation optimization, training framework optimization, machine learning and other work.
  • Hannah: I graduated from University of Manchester with a BACHELOR’s degree in electrical and electronic engineering and a master’s degree in data Science from University College London. I joined Huawei and have been working there for two years. As one of the pioneers of private cloud in China, Huawei has a lot of precipitation in the field of data science. Therefore, I had the opportunity to participate in multi-dimensional work, from algorithm research to platform development, POC project development, field modeling PK, etc., which deepened my understanding of data science in a short period of time.
  • Doctor Zeng: Doctor from a key university in China, huawei Special Offer in 2018. After working for five years after graduation, I returned to school to continue my master’s and Doctoral studies. During my master’s degree, I began to assist my supervisor in making projects. I have been in the IT industry for more than ten years, and I have rich experience in information system project development and management, as well as actual project experience in artificial intelligence. As the technical backbone of the laboratory, I have participated in several national, provincial and ministerial projects. The main research fields during the doctoral period are NLP and knowledge graph, and now I serve as the chief expert of an algorithm modeling team of Huawei.
  • Fan Ge: 20 years of ICT working experience, 4 years of signal processor development, 5 years of enterprise digital communication product research and development, 3 years of operating system architecture design, 3 years of big data analysis and research, 5 years of AI product planning. Currently, HE is the chief product management expert of Huawei Cloud ModelArts.

The body of the

MaiKeZhou

Knowledge cannot come from experience alone, but only from the invention of reason and from the comparison of observed facts — Einstein

As a software engineer who likes to understand principles from top to bottom, I did a lot of reading and hands-on work on each transition. My advice is based on their own actual situation, from the overall to specific technical books, read one book, don’t hurry.

The first book that I read is Nick “a brief history of artificial intelligence, the book almost entirely about the history of artificial intelligence, covering almost all areas of the discipline of artificial intelligence, including the origin of artificial intelligence, automatic theorem proving, expert system, neural network, natural language processing, genetic algorithm (ga), deep learning, reinforcement learning, super smart, philosophical problems and future trends, etc., Of course, it is not a hands-on programming book, but to give you a macro impression, suitable for AI product managers, Ctos to read.

If you feel the need to expand your overall understanding of technology, I recommend reading Artificial Intelligence (2nd Edition) by Stephen Lucci and Danny Kopec, which is kind of like a college textbook on ARTIFICIAL intelligence. It’s called the “encyclopedia of ARTIFICIAL intelligence.” The book covers a brief history of artificial intelligence, search method, knowledge search, game of logic of the search, artificial intelligence, knowledge representation, production system, expert system, machine learning and neural network, genetic algorithm (ga), natural language processing, automatic programming, robot technology, advanced computer game, the history and future of artificial intelligence and other topics.

After reviewing the global knowledge, it is suggested that you choose books based on your actual situation, such as Machine Learning by Zhou Zhihua, Deep Learning by Ian et al., Deep Learning by Hands-on Learning by Aston Zhang et al., TensorFlow by Zheng Zeyu et al. Google Framework for Deep Learning in Action (2nd edition), PyTorch Deep Learning by Vishnu Subramanian, these are good books, but there are many other great books that I won’t go into here. It depends more on your current state, whether you want to get your hands on the model quickly or understand the principles. It varies from person to person.

In addition to systematic reading and learning, WHAT I hope most is to get started coding and training models as soon as possible. IDE tools are necessary to start work. I am not used to IDE of public cloud, but I also want to use the powerful computing resources of public cloud, so I hope to have tools to help complete the linkage between local IDE and public cloud platform. I will tell you a case that has been realized — how to use PyCharm and ModelArts public cloud service for joint development to quickly and fully utilize cloud GPU computing resources.

I am connected to huawei Cloud public cloud AI platform. We actually use a PyCharm ToolKit to help establish the connection channel from the local PyCharm IDE to ModelArts. In this case, I use MXNet to realize the example of handwritten digital image recognition application, and quickly write the code locally. Publish to ModelArts public cloud to complete model training and model generation, and the generated model can be further deployed quickly (this step is not covered in this article). Before install the ToolKit, I need to install version 2019.2 (currently ToolKit adapter this version only) PyCharm, download address is: download.jetbrains.com/python/PyCh…

Note that if a higher version of PyCharm is already installed, you need to uninstall (automatically) the installed PyCharm first:

Download a tool PyCharm – ToolKit – PC – 2019.2 – HEC – 1.3.0. Zip, connection between local IDE and cloud link: www.jetbrains.com/PyCharm/dow…

Then go back to PyCharm IDE and open Settings:

Find Plugins and select a plug-in:

The following screen is displayed after the restart:

Then we need to go to Huawei cloud to declare the OBS secret key:

After the SMS verification code is successfully registered, save the CSV file to the local computer. Back to PyCharm IDE:

Note that you need to re-click the Edit Credential button to exit and see the check.

In this way, we have completed the docking of PyCharm IDE and ModelArts and moved on to the next step of actually training a model. First, download the hand writing data sets, a download link is as follows: modelarts-cnnorth1-market-dataset.obs.cn-north-1.myhuaweicloud.com/dataset-mar… Log in to huawei cloud and upload OBS.

After creating the folder, open the project in PyCharm and fill in the parameters. You can refer to the parameters filled in the ModelArts training model:

Click on “Run Training Job” and the lower right corner is the Training log returned to PyCharm from the public cloud:

After the training is completed, the training model is saved in the OBS of the public cloud, and you can choose to download it or do reasoning on the cloud.

Having a bunch of tools like this is a boon for those of us who actually write code.

Mr Qiu

“The queen of A Certain county; Quite was at peace with the county people; Day to know everywhere safe; Early harvest medium ripe; Feel the wind and rain as scheduled; Late rice is also expected; But is the force cotton for grazing; There are days to come.” — Song Wen Tianxiang and Hongdu Rui Ming Yunyan Book

Before I introduce how I made the transition into AI, I would like to introduce my personal experience. It has been about 7 years since I got in touch with AI. I can divide it into 3 periods :(1) study in school; (2) Internship transition; (3) Working growth stage. I learned a wide range of subjects at school, mainly in artificial intelligence, high-frequency electronic circuit, traditional image algorithm, embedded system, etc. At this stage, I achieved some good competition results by promoting learning through competition. During the internship, I was fortunate to participate in the module of “Search for pictures” of a key project, and accumulated a lot of practical experience in image and artificial intelligence. After graduation, I started to work, mainly focusing on the research and practice of computer vision, algorithm transplantation optimization, deep learning framework optimization, machine learning, etc., and gained certain experience in full-stack AI. I have a feeling that although I am not qualified, I have a clear goal, I have been working hard, and I have met a lot of people worth thanking, so I have the opportunity to introduce my experience here.

In 2012, in the annual ImageNet image recognition competition, AlexNet of CNN surpassed the second place in the classification performance of crushing (SVM method), and from then on deep learning began to attract the attention of researchers. However, it really attracted the industry to invest in deep learning on a large scale from 2015 to 2016. At this time, Jia Yangqing opened the deep learning framework Caffe, the first open source version of Google’s ARTIFICIAL intelligence engine TensorFlow, and Caffe used to be famous for its excellent code architecture. Convenient development interface and other advantages by the majority of researchers and engineers praise. Do AI development early, many tools are not perfect, need to make the wheel, the wheels here mainly includes offline training need to develop image annotation tool, need to modify the deep learning framework implements the forward and backward calculation, related operator to tune the convolution between feature visualization model, such as online deployment of the need to achieve operator FP16 and int8 calculation, Need to write your own CUDA code etc.

Now, developers entering the AI space don’t have to make those wheels anymore, which is really nice.

Image annotation tool LabelImg and other open source, can be directly taken over to use; TensorFlow has been iterated to make it easier for developers to use, providing excellent examples, visualization tools, and reasoning deployment tools. With the rise of PyTorch and its ability to score with TensorFlow, developers have more options; Nvidia has launched its GPU-based inference acceleration tool TensorRT, which is free to use, and Huawei has launched its inference acceleration and framework MindSpore, which is based on its own chip Da Vinci, to enrich developers’ choices with higher computing power density.

It’s a good time for AI. It’s a good time for AI developers. I think it’s a good time to start making the transition to AI.

Looking back at the development of AI over the years and the pit we have stepped on, I will extract the method of transformation from traditional software development to AI development project from my personal point of view, hoping to be helpful to everyone. These can be summarized into three parts: (1) start running, (2) master the principle, and (3) deploy the reasoning.

1. Start running:

This stage is to learn through the excellent existing open source projects, here I recommend Darknet YOLOv3, the project address github.com/pjreddie/da… . The project framework is based on C language, the framework is clear and easy to debug. The first thing beginners need to do after downloading the project code is run the project and get a general idea of how the AI works and what problems it can solve by tracking the flow of data through the framework. At the same time, in the process of running and building the environment, we can quickly master how GPU and deep learning framework work together, and what dependence the deep learning framework needs in the process of running.

2. Master the principle:

As an excellent and classic end-to-end detection algorithm, YOLO is recommended for learning. Since its launch, YOLO has been iterated to the fourth version. You can start with the first version of YOLOv1 and move on to the fourth version of YOLOv4 to learn how YOLO has evolved, what improvements have been made, and why the improvements have improved performance. There are plenty of great blogs out there that developers can search and read on their own. Combined with the start up phase of the code debugging, I believe that progress will be faster. After studying the evolution process of YOLO, if you are interested, you can study the detection algorithm of R-CNN series. R-cnn series algorithm is different from YOLO in that it has high recall rate and high accuracy, but it is time-consuming. Therefore, YOLO is generally used in the industry.

3. Deploy reasoning:

Reasoning deployment is to deploy the model trained by deep learning framework to PC or end test device for reasoning to solve practical business problems. This process mainly includes model performance tuning, GPU or D chip adaptation, and business logic implementation. The model performance tuning is mainly to quantify the original FP32 of model inference into FP16 or INT8, so as to realize inference acceleration and real-time inference. Here we recommend Nvidia’s TensorRT and Huawei’s D chip acceleration module, which will perform higher-order optimization on the model. In addition to quantization, there will also be operator fusion and Kennel optimization between network layers of the model, etc. Specific information can be searched on the corresponding official website. After model optimization is completed, online reasoning of the model can be completed through corresponding inference engine and business logic realization, so as to complete real industrialization and solve practical problems, which is also the part where AI really generates value for the society.

After learning these three parts, I believe you will gradually become a senior algorithm engineer in the AI field.

Hannah

Now that I have set foot on this road, then, nothing should prevent me from going on this road. – Kant

Hello, everyone. In this article’s group of five, I am a little white at work. As a little white person without much social experience, I spent most of my life in school. I will mainly introduce my school experience here.

When I was an undergraduate, I chose electronic and electrical engineering as my major. At that time, I had no idea about my major or my job, but I just thought that science students should choose a “hot” major. So I chose the most hot major from the long list, and then I went to Manchester happily with my backpack on my back to start my university career.

After the semester began, it was really different from what I expected. This kind of engineering major requires students’ hands-on experiment ability. When I saw a pile of experimental equipment in my first experiment class, I was completely confused. At that time, the scene was quite scary for me. The Indian boy in the same group listened to the requirements of the experiment teacher and clicked his operation. I stood beside him and even forgot the important thing of recording his operation and going back to study. Only after he had the results of his experiment did I ask carefully what the operation was all about. My brother was warm-hearted and explained it to me in quick, Indian-accented English. But I learned English listening, only heard British and American, for Indian English is really strange, I had to cheeky little brother again. At the third time, he finally lost patience and pointed to a section of the textbook for me to read. I realized that he had been saying it all along, but I hadn’t understood it, including the time he pointed and read it. Therefore, I spent the first experiment class in the process of checking the names of various instruments and the high-speed curry English explanation of the Indian boy in the same group. I am grateful that the team did not give up on me and let me pass my freshman year smoothly.

When I chose my design direction after graduation, I had a brainwave and chose a noise-reducing earphone project related to DSP (digital signal processing). Here’s how it works:

Generally, there is a small microphone inside the active noise-cancelling earphones to collect the external sound. Through calculation, a completely opposite sound wave can be generated and then superimposed with the external noise to reduce the noise.

The whole project is just in accordance with the operation of the teacher’s teaching materials and reference books, to achieve a FIR-LMS algorithm filter. Although the algorithm is simple, but the implementation effect is quite good. I made a small set of microphone hardware to demonstrate the results. I played white noise with my mobile phone beside the microphone and talked at the same time. Finally, the sound from the stereo had a partial noise reduction effect.

When doing this project, THE biggest feeling is that this algorithm can not only get some achievements in scientific research, but also bring a lot of convenience to us ordinary people when it is applied to life.

At that time, I was thinking about how I could be exposed to more of this technology and even participate in the implementation of this technology. At this time, a cartoon told me the answer. Baymax, the advanced ai from Big Hero 6, is a near-all-powerful family doctor. This is the power of AI. Among the offers of several graduate students at that time, I finally chose data science as my graduate direction.

It was not easy at the beginning of my postgraduate study. It was not very smooth to transfer from electronic and electrical engineering to data science. Data science this major is statistic department and computer science department cooperate do, that is to say, this major is taller to statistic and computer science requirement. At that time, the teacher also listed a long list of references, PRML (pattern Recognition and Machine Learning) and BRML (Bayesian Reasoning and Machine Learning) impressed. If you are interested, you can also go and have a look. There are also good books in China. I recommend Li Hang’s Statistical Learning Methods and Zhou Zhihua’s Machine Learning.

This major still attaches great importance to our practical modeling skills, and more than 30% of the scores of almost all courses require us to do modeling. One of them was applied machine learning, which required us to compete on Kaggle and use our rankings to score our experiments. Within one month, our group participated in three projects, among which we got the top 4%. The projects on Kaggle are very suitable for beginners to practice. After learning a lot of theoretical knowledge, beginners will often be unfamiliar with a bunch of mathematical formulas and do not know their practical effect. Competitions on Kaggle give beginners a chance to get in touch with real data and use it for modeling. They can participate in these competitions and learn whether the models they are currently learning can rank high in solving real problems and whether they are really better solutions based on their own results. This gave me a deeper understanding of the theories I learned in class in a short period of time. Even if the actual meaning of the data is not clear, even if the meaning of the data is encrypted, some machine learning techniques can be used to model and predict the results. Titanic project in Kaggle is also the first choice for many teachers to give students the introduction to data analysis and machine learning. Those who are interested in it can have a practical experience.

There are different requirements and applications for models in academia and industry. In order to better understand the actual modeling process, I chose a cooperative project between the university and the company after graduation. My graduation project was my first internship in data science, making a simple recommendation for similar singers and a map of the best cities for singers to tour in the UK for Warner Music company. In this project, I got in touch with real customer data and worked with companies to try to mine more connections in the data. At the same time, I was exploring more possibilities and entered another NLP contest on Kaggle to judge the similarity of two short texts. When those programs ended, so did my graduate years.

After returning to China, I took a break and joined Huawei. In Huawei, my first big data-related project is PyTorch mass acceleration training, which is similar to MoXing on Huawei cloud, but it is based on PyTorch optimization. In this project, I had the opportunity to get in touch with some basic concepts in deep learning computing, and had an understanding of the communication between Gpus. Later, in order to add more models to the platform, I had the opportunity to take a closer look at the model’s architecture and resurrect the popular MobileNet effect. Then we learned and successfully embedded r-CNN series of target detection algorithms in our products.

After this deep learning experience, I invested in the machine learning platform and basic algorithm implementation of our products. I embedded LightGBM algorithm in Huawei cloud, and did some optimization work on our machine learning platform.

Then, I led the algorithm model POC project of a national unit and the algorithm co-creation project of credit card center of a major domestic bank. Exposure to real data again, and modeling against the data. Both projects are risk control projects, and modeling in real data is really interesting. Every time you model with data from a different industry, it’s like you get a whole new understanding of another industry. During this period, I also had the opportunity to introduce some popular science knowledge of machine learning to VIP customers in English on foreign forums.

In my two short years on the job, I tried every job in my field. I still think my major is great. It opened up a new world for me.

Doctor Zeng

An inch of time an inch of gold, an inch of gold can not buy an inch of time — zengguang Xian Wen

As a doctor in computer science with many years of practical work experience, I will talk directly about my understanding of the modeling process and the corresponding practical cases.

For IT practitioners, most of the project managers and engineers engaged in the development of traditional information system projects may be attracted by this fashionable name, but they do not know how to start to build a lofty AI project, just like the ancient Yang Zi, they only know to squat down and cry when they meet the wrong road. Fortunately, this is the most beautiful era, we no longer have to look at thousands of roads do not know how to move, a large number of scientists and engineers wearing the halo of wisdom, has already laid a solid foundation for us, there are many AI modeling platforms even amateurs can quickly get started.

When it comes to machine learning, many people may balk at it, believing that they need to master very advanced mathematical knowledge and all kinds of obscure theories. Unfortunately, a solid knowledge of mathematics and theories of machine learning is required for AI professionals, but not necessarily for building an AI project. Here I’ll start with a simple scenario to help you understand the process of building a typical machine learning project.

The so-called life is for the benefit of all, then we take the scene in the financial industry. Finance is probably an area of great interest to everyone, and even I, who let it slip, know a few financial terms, such as credit card fraud. Since the first thing that comes to mind is credit card application fraud, let’s talk about the AI construction process using credit card application fraud detection as an example.

All things are difficult at the beginning, then in the middle, and finally at the end. For many people, AI projects seem to feel this way, always elusive. Some people say that the three elements of AI are “data”, “computing power” and “algorithm”, but I think there is another indispensable element — “business”. Different from academic research, engineering project is usually for a specific business scenarios and specific business objectives, we are in the process of building AI project at ordinary times, always habitually to the whole process is divided into the data processing, characteristics of the engineering, model training and model inference and model deployment process, but in fact, the analysis and understanding of business throughout the project. One senior said that with his experience in AI project development, he understood the business, had enough data, and found the right algorithm, the project was in the clear, of which understanding the business is the crucial step, I deeply believe.

Since business understanding will continue throughout, let’s talk about business understanding first. Simply put, business understanding is understanding what we are trying to do (business knowledge), where it is (available resources), and what we are trying to achieve (business goals). For credit card fraud detection project, in addition to the preliminary works such as project preparation, resources to prepare, you first need to do is to understand the credit card business, know what is credit card fraud, why it’s easy to have a credit card application fraud, what we have available data resources, as well as credit card fraud detection system goals to achieve, and so on, Then I think about how to do fraud detection on credit card applications. The process of business understanding often requires the deep involvement of business experts and close collaboration between algorithm modelers and business experts.

Once the business requirements are understood, the acquisition, analysis, and processing of data becomes an important step that cannot be bypassed. Data acquisition is relatively easy to understand, simple point said is on the premise of legal compliance, collected as much as possible with the business problems closely related data, such as credit card applicant to fill out personal information in the process of apply for a credit card, the pedestrian credit report, third party credit analysis report and so on, all can be used as credit card application fraud detection system of the input data. The main purpose of data analysis is to understand data and extract useful information from data, which is another big category with a variety of tools and means, even throughout the entire AI project construction process.

Generally speaking, the data obtained from different channels are of varying quality and full of a large number of redundant, repeated, missing, abnormal and inconsistent data, which is difficult to be directly used in the construction process of AI model. The determination of redundant data needs to be combined with business analysis, correlation analysis and other means, usually including business invalid data, or data highly related to other data, such data and repeated data processing method is very simple, simply delete; There are many ways to deal with missing data, including simple deletion, special value filling, related data derivation and other conventional methods, as well as hot card filling, clustering filling, and high-order methods based on simple machine learning model prediction. The processing of outliers should be analyzed by box graph, three Sigma criterion, DBSCAN clustering and isolated forest, etc. After identifying outliers in the data, the missing values should be processed. Inconsistent data mainly refers to the data with the same meaning but different representation methods, such as mixed case, irregular date format, irregular address and inconsistent unit, etc. Such data only needs to be unified with data rules.

Once the data is organized, it is often necessary to slice the data set. When the amount of data is large, it can be divided into training set, validation set and test set, and directly use the training set to train the model, use the validation set to determine the best model parameters, and use the test set to evaluate the model performance. When the amount of data is small, the model parameters are determined by cross validation instead of special validation set to ensure that the model is fully trained. The total amount of fraud detection data of credit card application is usually very large, but few of them actually have fraud, that is, there is a very serious data imbalance. Therefore, in the process of data set segmentation, it is necessary to ensure that fraud data can enter each data set in a certain proportion. The screenshot below contains the platform of the data segmentation operator.

After data processing is completed, it enters the stage of complex and delicate feature engineering. In traditional machine learning projects, data processing and feature engineering have a very high status, and some people even put forward that data and feature determine the upper limit of machine learning, and algorithms and models are just approaching this upper limit. In our practice, the entire process of data processing and feature engineering takes up at least 70 percent of the development effort of the entire project, and often involves going back and forth to polish features during the modeling process. The image below is a screenshot of the feature engineering workflow.

Finally, the data for the modeling is ready, and we can start to do some fancy modeling. I prefer to call this process model engineering, which includes model selection, model training, model evaluation, and model reasoning. The process of model selection usually involves a lot of basic machine learning skills, as well as a deep understanding of data, business, and possibly some modeling experience (as well as the experience of others).

As we know, machine learning tasks can usually be divided into supervised, unsupervised or semi-supervised tasks based on whether the data set has labels, and classified problems and regression problems based on whether the predicted data is discrete or continuous. The easiest way to deal with fraud detection on credit card applications is to treat it as a supervised dichotomy, where we only need to determine whether a user-initiated application is fraudulent or not. Now we use LGBM, XGBoost, RF, etc., as shown in the figure below.

After the model is selected, the process of model training, reasoning and evaluation begins. The training process of the model is to feed the prepared training data to the model so that the model can learn the rules and rules contained in the data in the form of parameters. Usually, during this process, you set up all kinds of hyperparameters that the model needs. There are various open source machine learning libraries and easy-to-use machine learning platforms everywhere. In most cases, we don’t need to build a machine learning model from scratch. We can simply switch around various open source machine learning packages or drag and drop operators on the machine learning platform. The process of model reasoning is well understood: feed the data from the test set into the trained model and let the model predict the result. For example, for fraud detection of credit card application, the data to be predicted is thrown into the model, and the prediction conclusion corresponding to each data is given by the model. The model evaluation process is to evaluate the predicted results through various indicators to measure the gap between the predicted results and the real results. For classification problems, common measures include accuracy rate, accuracy rate, recall rate, F1 value, etc. For regression problems, common measures include MAE, MSE, etc.

Screenshot of the evaluation operator result display interface of ModelArts Miner platform

The process of model training, reasoning and evaluation, and even data processing and feature engineering, may take a lot of tweaking and hammering, but with a lot of hard work, the stick can be turned into a needle, and you’ll eventually get what you want. Well, let’s say you’ve got what you want, and your model can predict, to some extent, whether a credit card application is fraudulent. At this point, the dawn finally lit up the eastern sky, the horizon also appeared a ray of fish-belly white. Now all you have to do is make the dawn come harder, deploy your models to your business systems, and let your data happily swim in the AI pipes, emitting dazzling fireworks. After deploying to a business system, you need to constantly review your results and adjust and optimize your entire model based on the latest data. But at this time, is not my article can help you, you spread the magnificent wings, to breakdown the sky, enjoy flying. The following chart is a screenshot of the whole process of credit card fraud detection based on the workflow mode of Huawei cloud AI platform.

The elder brother of the sail

Heaven line, the gentleman to unremitting self-improvement. Zhouyi · Gan

As a senior product manager, I have been engaged in AI platform product design for many years after my transition to AI. Let me talk about my understanding of AI products. For the public perception of AI development, we often think of algorithm development. In fact, for a commercial AI development process, it will involve many different aspects. It is not only algorithm development, but also calculation power, data, iterative optimization and other links. For the commercialization of AI, Ctos will consider the investment of resources in three aspects in a balanced way, not only in IT infrastructure implementation but also in engineers:

1. Computing power — > Universal benefits;

2. Data –> model;

3. Algorithm –> Landing; Each link involves different human input, generally involving IT engineers, data scientists and application engineers.

To elaborate:

1. Computing power — > Pratt & Whitney (for IT engineers)

For AI development, how to make good use of computing power resources and improve resource utilization rate is a particularly important work at present. Engineers often algorithm is not good at these things, need to have professional engineer system level to help is to implement, and from the cluster of the structures, operations, firmware distribution operations, the operator optimization, tuning, and the framework of a lot of maintenance, etc., in addition to the resource management between different departments is IT engineers need to focus on content. For simple development, it’s ok to build your own open source infrastructure, but when it comes to large-scale commercialization, it’s important to have a good platform.

2. Data –> Models (for data scientists)

For AI modeling, it is essentially a process of generating models based on data, so generally speaking, the initial demo can be produced in a few days, but when it comes to the actual implementation, it will take at least months or even a year. This process involves two types of massive input. One type of input is to acquire training data from end-to-end data collection and annotation, and the other type of input is to continuously iterate model optimization, which involves a large number of domain skills.

Training data: scenes oriented to general classes will involve a lot of data annotation work, which will cost a lot of annotation manpower. For general scenes, ordinary personnel can annotate, but for complex scenes, professionals are required to annotate, such as medical image scenes, ordinary people can not annotate. Therefore, it is particularly important to have a very friendly algorithm ability for automatic annotation and data difficult cases for annotation in professional fields and a large amount of data.

Iterative optimization: model in the process of modeling business scenario, data scientists during the process of modeling, the use of more mature papers and algorithm, around specific business scenarios, combined with the data, to selectively modeling, and changes with different environmental cases, constant adjustment algorithm design and combinatorial optimization, so as to achieve the best effect of the algorithm. There is a lot of scene understanding involved in this process. For developing models, it is helpful to have a good case base to quickly find scenarioized precipitation domain knowledge.

3. Algorithm –> Landing (for application engineers)

In the past, we think that algorithm landing is a similar integration process of software development, in fact, the application of AI class is not so simple. In general very general class scenarios, integrate existing models to do some relatively mature recognition, such as speech recognition. However, the suitability of commercial scenarios is not good enough. Although the industry is discussing the possibility of identification of everything, it is not satisfactory. Therefore, for the actual implementation of the algorithm, it is necessary to constantly optimize based on the actual situation. Because of the iterative optimization process, application engineers must focus on the business pathways and convenience of deployment and training in thinking scenarios. Generally, data scientists and algorithm engineers are relatively scarce, so it is not feasible to send data scientists and algorithm engineers to each site for tuning implementation. A large number of field implementation work must be completed by a large number of application engineers. It is a rare problem in software engineering to optimize the model for application engineers and to iterate and adjust the precision of the model.

This often requires a systematic platform to support rapid iterations of business scenarios, thus improving the efficiency of application engineers.

From nearly 20 years of work experience, commercial implementation of an AI business is actually the most complex system engineering. It is not just a single algorithm development, but the skills required to land around commercial AI are extraordinary.

1, from the perspective of the IT engineer, he needs to do different focus resources efficiency and cost considerations, around the underlying resource management, system architecture design, distributed optimization, management of resource scheduling, IT construction design, end – side – cloud deployment and the corresponding design, easy maintenance and so on made which can be a very good support algorithm is applied to the ground;

2, from the perspective of a data scientist, he just want to Know how to develop an algorithm and parameter optimization ability, but also to understanding the industry scene, with industry Know – how, according to the environment and the problem of data, iteration and modeling, constantly adjust algorithm to deal with environmental changes, to achieve the corresponding effect, will need to pay attention to during a lot of ethical and moral, Including security compliance, model security and interpretability issues;

3. From the perspective of application engineer, he is not only an application integration, but also needs to understand the uncertainty of AI application, select appropriate scene constraints, and adapt to the ability of existing models according to the actual situation. In addition, I have the ability to conduct training and iteration of the model independently, through the pipeline preset by data scientists, and optimize and iterate on the spot by myself. So as to achieve the landing of the real scene.

Summary: For many commercial companies, they focus more on the commercial scenarios of AI, and tend to ignore the importance of AI platform. But often for ctos/CIOs of forward-thinking commercial companies, they will consider the AI platform choice when AI is commercially launched. The original intention of Huawei Cloud ModelArts platform is also born from the difficulties of AI commercial landing. Combined with Huawei’s own Centrum chip system, ModelArts is constantly committed to the universal benefits of AI.

Afterword.

After decades of development, computer science has become a mature discipline, the current university computer department organizational structure chart, each computer department mostly has three groups of people: theory, system and AI (artificial intelligence). Twenty years ago, there was a saying in American computer circles: theory and system people looked down on each other, but they looked down on AI people at the same time. AI has been hot these years, but it was also the underdog. Philosophy once gave birth to science, but as soon as the question was settled, it became a separate science.

As Allen, who has said, the history of AI is a saga of antagonistic struggle issues, such as analog and digital, serial and parallel, replace and enhance, grammar and semantics, mechanistic and skopos theory, biology and vitalism, engineering and science, symbols and continuous, logic and psychology, etc., in each issue of further can be divided into sub topics, Such as in logic and psychology and theorem proof and problem solving, there is room for the development of controversy.

I believe that this trend of struggle will still exist for a long time. Only by adapting to such struggle and progress, and constantly enhancing our technology and meeting the challenge of new technology, can we keep our career going forward. Therefore, you can take advantage of the package training programs provided by big companies in the industry to continuously enhance your technical depth. For example, Huawei Cloud designed a high-quality course 2020 Huawei Cloud AI Training Camp for all AI developers. The content includes the basic knowledge of image classification, object detection, image segmentation, face recognition, OCR, video analysis, natural language processing and speech recognition, as well as the introduction and practice of classical data sets and classical algorithms. Each chapter of the course is a practical case carefully created by Huawei cloud AI experts, the whole process covers model training, testing, evaluation, with code explanation and homework, to help you master the model development ability of eight popular AI fields, transition to become an AI developer.

Zhou Mingyao, member of Jiusan Society, graduated from Zhejiang University in 2004, master of engineering. Currently, he is the r&d director of Huawei cloud AI products. He is the author of “Big Talk Java Performance Optimization”, “Deep Understanding of JVM&G1 GC”, “Technical Leadership – How to Lead a Software R&D Team”, “Programmer Training Record” and so on. I started my career as a software engineer, and later became a distributed technology engineer and a big data technology engineer. I began to contact AI technology in 2016. Michael_tec.

Click to follow, the first time to learn about Huawei cloud fresh technology ~