E qualification sponsored by Japan Deep Learning Association JDLA certification program Study-AI Co., Ltd., which is known for providing certificate /? Id = certificate_No04), has set up an AI implementation certification executive committee, and as a qualification to acquire the skills to challenge the E qualification, "AI implementation Test ”has started. I took the exam: cherry_blossom: Passed successfully: cherry_blossom: Yes. Therefore, I would like to dump information such as test preparations that I actually worked on before my memory faded.

Actually received pass notification email

It seems that the problem with Python is that you made a careless mistake. .. ..

This qualification is an article by @ kazzy0099 who took up Past article [[Updated from time to time] Summary of qualifications related to data science and machine learning]( I found it at https://qiita.com/kazzy0099/items/d423491bda6fedbce196) and took the exam. It will be updated from time to time. I look forward to future updates.

Motivation for taking the exam

I took the exam with the following motives.

Because I am interested in and studying machine learning and deep learning
Because I want to take the E qualification in the future
Because I thought that taking the qualification that put the spotlight on the implementation would lead to strengthening the implementation ability
Because my hobby is getting qualifications, etc ...

What is an AI implementation test?

Details can be found at the links below.

A sketch of the position of the test was posted on http://kentei.ai/.

There is a possibility that grades other than A grade will be formulated in the future, but this time only A grade was recruited. In addition, this test is a test that can be taken on your home PC . A total of 60 questions will be given, 20 for each field of mathematics, programming, and deep learning. The test time is 60 minutes, so you can't afford to think twice. Regarding the passing score, it seems that the total score must reach the passing score announced for each test. The examination fee is ¥ 3,500. Recently, there are many qualification exams with high examination fees, but the price is reasonable. (The author's view)

: warning: Passed AI implementation test A grade = Not qualified to take E qualification. As a separate test eligibility requirement, you must attend JDLA Certified Training.

2020/3/30 postscript About pass rate and pass score

Report: AI implementation test [A grade] 1st implementation result, the pass rate is about 80%. I can't say for sure because the passing score has not been announced, but as far as the average score rate is concerned, you may want more than 80%.

Test range

Test question sample also serves as the exam question range. There is also a reference to the composition ratio of all 60 questions.

Test preparation

I would like to introduce the teaching materials that were actually used for exam preparation. The study period was intermittent, so I can't say exactly, but if you have some knowledge of neural networks and deep learning, I think you can do two laps of the official teaching material videos in about two weeks.

: warning: If you want to pass from zero knowledge, you need a certain period of study. (Especially mathematics) : warning: I'm a science student, and NN and Deep Learning are a little bit of knowledge in my graduate classes and research.

Official teaching materials

It is published in AI Implementation Test Official Teaching Material. This time it was ¥ 3,000. [^ Text] for a limited time. It is a teaching material that can be purchased only by examinees without commentary by the instructor (with video commentary + source). It seems that the official version with video commentary will be released in the future. The price will be ¥ 45,000. It seems that the video of this teaching material can be viewed only until June 30th.

: warning: This teaching material is mainly used as a teaching material for AI and programming in the examination range. Separate measures by self-learning are required for areas not included in mathematics and official teaching materials.

Official teaching material outline

From http://kentei.ai/introduction/howto

★ Bambi β-Taurus (Section 1 Basics) If you want to proceed from the basics of Python and the operation of your computer, start here. Learn how to use Numpy and Matplotlib, which are libraries required for drawing functions and numerical calculations that are essential for neural networks.

★ Bambi β-Gemini (Section 2 Basics) The basics of forward propagation of neural networks start here.

★ Bambi β-Cancer 1 (Section 3 application) Back propagation of neural networks. Beginners will graduate if they step up to this point.

★ Bambi β-Cancer 2 (Section 3 application) Chain rule of neural networks. If you can understand this, you are an intermediate person.

★ Bambi β-Leo (Section 4 application) An update-type implementation that introduces an intermediate layer. The source code will also be distributed, so please actually move your hands to "learn" it. If you can do this, you will be a very talented person.

★ Bambi β-Virgo 1 (Section 4 application) Implement deep learning with handwritten character data. If possible, intermediate students will graduate. You can reach a level where you can learn deep learning even if you study by yourself.

★ Bambi β-Virgo2 (Section 5 application) ← The content of Section 5 is out of the scope of the test, but it is interesting so it is worth working on Implement deep learning with handwritten character data. If possible, intermediate students will graduate. You can reach a level where you can learn deep learning even if you study by yourself. Using currency exchange data as an example, we will implement it in five languages and frameworks. (Chainer, NumPy, PyTorch, TensorFlow, Keras) Forex data has also introduced the concept of dealing with time series called RNN, and it goes beyond the world of Bambi to the advanced level.

Sample problem

I also solved the problem of Test question sample on the official website. There was no answer, so I created a rough solution for reference.

Input layer and output layer: (a) Input
Forward Propagation Calculation: Question 1 Color Code, Question 2 27
Matrix multiplication: (H) 2, 8
Introduction of bias term: Option 1

programming

Numpy 1st question: Option 1 (However, there is a typographical error in which w and x are interchanged)
Numpy 2nd question: (c)
Pandas: read_csv
scikit-learn: Divided into train data and test data

Math

Sequence and matrix 1st question: Problem statement mistake?
Sequence and matrix 2nd question: A: 10 B: 22
Sequence and matrix 3rd question: A: 6 B: 10 C: 14

Commercial books

Introduction to self-made neural network (written by Tariq Rashid, translated by Hiroyuki Shinno)

This is the Japanese version of the book Make Your Own Neural Network.

makenn

When I read it again after doing the official teaching materials, I found that this book was enthusiastic about the easy-to-understand explanation of theory and logic, and while keeping the rigor, the neural network was explained in a way that was easy for the reader to understand. I realized once again that it was an excellent book that explained. This book is a Japanese translation, so I would like to read the original.

Impressions of actually taking the exam

In the actual exam, 20 questions will be given in the order of mathematics, programming, and AI. The first mathematics mainly requires knowledge of linear algebra. At first, it may be annoying, but since it is a test that does not have much time to think and solve 60 problems in 60 minutes, it is necessary to choose to put off or discard the calculation problems that may take time. Will be. In programming problems, there was scikit-learn in the range, so the questions were given at a fairly high rate. The AI part was a question that was roughly the same as the official teaching materials. (I love the gradient descent method) It may be because it is a qualification that is conscious of E test preparation, but as mentioned on the test page, I got a strong impression that it is a test that emphasizes implementation and reasoning. After that, it is better to have a piece of paper and a pen so that you can write the calculation problem by hand and solve it. I have taken many qualification exams so far, but this is the first time I have taken the exam at home. It may be the perfect format for today's world.

An article by @ PoppeDwarf Impressions and memorandum of AI implementation test was posted. The author of this article also commented on the same thing as I did.

: warning: An email requesting you to log in to the test system in advance and check the operation will be sent to you a few days before the test date. We recommend that you log in and check if it works properly in advance. In addition, prior to that, suddenly, Study-AI will send you a learning content usage page and login information [^ studyai], which is the login information for the test system. At this time, I thought it was spam and hesitated to open it.

Summary

We have summarized the AI implementation test class A. I think this exam was very meaningful in that I was able to re-learn about Neural Network this time and reconfirm that my understanding was weak. In the future, I would like to continue to learn the latest topics such as machine learning, deep learning, quantum computers whose theory is difficult to understand, and quantum machine learning. This article is the 20th article posted since August last year. I would like to continue to actively disseminate information.

2020/4/13 postscript The pass certificate has arrived

I received a pass certificate. It is said that he will be awarded the title of Deep Learning Implementer. It may be the first time in my life that I got a qualification and got a title.

合格証.png

Two types of pass logos were also attached, so I decided to replace the Qiita icon with the pass logo image from today.

Future goals

I would like to take the exam when the S class appears.
Rereading Deep Learning from scratch
Completion of Deep Learning 2 made from scratch
Relearn linear algebra (mainly only the parts required for machine learning)
Master of model implementation with TensorFlow, PyTorch
Secure E qualification exam requirements
Make Your Own Neural Network (also for studying English)
When a book to make a quantum computer is published, I will try to make it etc ...

[^ text]: Limited teaching materials for the second time are also sold for ¥ 3,000. This is also available for purchase only by test entrants, and the viewing period is also limited.

[^ studyai]: The subject of the email is "Notice from content learning AI (Study-AI)".

I passed the 1st AI implementation test [A grade], so I tried various things