How Should Engineers Learn Machine Learning Algorithms

1. Preface - Programmers are a profession that updates knowledge more rapidly. The speed of updating knowledge in this profession sometimes exceeds your learning speed. Most people are constantly updating their knowledge system. Here I want to show readers the importance of machine learning (including in-depth learning), as well as common knowledge. As an improvement of personal ability, the author thinks that this is a procedural ape/Yuan, which needs to master a thinking skill. In fact, all the students who have worked in Internet companies should have this feeling. We are transforming from IT to DT. Algorithms are becoming more and more important in applications.

How Should Engineers Learn Machine Learning Algorithms

In the transition to DT, we also need to master the core of DT - algorithm. External big data is in full swing, but for the industry, these things are just noisy concepts. Simple big data is not very meaningful. Data is like fuel for engine operation, and algorithm is engine. Only with engine data can its value be brought into play. Just as oil has been sleeping in the ground for millions of years, nobody realized its value before the engine came into being, but after the engine came into being, oil became as precious as gold. So we must master the knowledge of algorithm in order to better control DT.

In practice, however, most of the so-called algorithm engineers should be called "algorithm application engineers". Few algorithm engineers can produce their own algorithms. I'm sure someone wants to make bricks, but it's true. For example, some classification problems are solved by logistic regression, SVM and other commonly used algorithms. For some complex problems, such as image classification using CNN, slightly more complex NLP problems using RNN. Many algorithmic engineers use most of their daily work in data processing, training data set selection, verification data set selection and so on, constantly adjusting parameters to achieve better results. This is not to slander those hard-working algorithmic classmates, I also engaged in machine learning when I was a graduate student. After work, I first engaged in engineering and mixed algorithmic work, such as classification problems, category mapping (mapping the goods of A website to the category of B website), and then engaged in NLP (slotting based on RNN and CRF)... And then we went on to work on the project. What I want to say is that the algorithm is not as complex as we think, but it is actually far more complex than we know.

Why is the algorithm not complex but complex? In fact, this is not contradictory. It's not complicated because we don't have to fully grasp the complex formula derivation behind it. We just need to know the physical meaning behind it, the problem that the response algorithm or model is suitable for processing, and how to adjust the relevant parameters. It's complicated because the deduction behind it is really troublesome. Take the commonly used SVM algorithm for example, when reading, it's hard to deduce successfully after a week according to PPT deduction (which also reflects that my mathematical foundation is a bit slaggy). Continuing to talk about complexity, to master the core of the corresponding algorithm very skillfully, we need to have a thorough understanding of the important mathematical knowledge such as linear algebra, probability, calculus and so on. Here we have to say that our college mathematics education is not thorough in linear algebra, probability and calculus. We only learned knowledge points but failed to clarify the underlying physical meaning. For example, the physical meaning of matrix correlation arithmetic is only realized in the actual algorithm. Mathematics should really make computers old. As far as teachers are concerned. There is a little bit of talk, but for programmers, I think mathematics knowledge must be solid, especially for students in schools, other courses can not be too attentive, but mathematics-related must be firmly grasped, to be attentive, especially to grasp the underlying physical significance.

Why should we master the algorithm? First of all, it feels that it is because of the change of environment, because the algorithm is becoming more and more important, such as the news before, the transformation of Google: turning 25,000 engineers into machine learning experts, Google has already been in the forefront, and it also shows that machine learning is becoming more and more important. Nowadays, many products, engineering only packages algorithms. The core of many products is algorithms, such as droplets, Uber.

Secondly, the engineer knows the arithmetic. It's a bit like a hooligan who knows martial arts. Nobody can stop it. On the one hand, to increase their core competitiveness, on the other hand, it is more convenient for team communication. Many algorithmic engineers lack the ability of architecture and system. If the engineer knows the algorithm, he can communicate with algorithmic classmates smoothly, which is helpful to the realization of the product or project.

Finally, for school students, job hunters, graduates or new students, mastering the basic knowledge of machine learning can increase your core competitiveness and make it easier to stand out in the face of competition (especially in interviews, there is no need for practical experience, you can speak clearly and very well).

In a word, I feel that learning machine learning algorithms is an important step for engineers to improve their core competitiveness. Then the author further elaborates what related algorithms should be mastered by engineers and how to learn them.

### 2. Basic knowledge But from personal experience, many graduates are interested in linear algebra, especially moments.

Please read the Chinese version for details.