Machine learning and learning ushered in the bottleneck period, talent shortage in the next 3 to 5 years

Nowadays, artificial intelligence is exploding, and more and more people are learning machine learning. However, Xiao Bian believes that we must not blindly learn, and the result of drifting will be eliminated by the times. Xiaobian has compiled some of the most in-depth machine learning talents in the next 3 to 5 years. We can choose the technology that suits us in this direction.

Now that I am in the industry, let me talk about what kind of machine learning talents the industry needs in the next few years. Not to mention the academic world is mainly because most people will not engage in research, but will struggle in the field of application. In comparison, the demand for talent in the industry is more conservative, which is different from the academic world. This is limited by many objective factors, such as hardware computing power, data security, algorithm stability, and labor costs.

This answer may be more suitable for two types of people:
1. Student friends who are reading

2. A friend who wants to switch to machine learning soon after work. The particularly powerful technology of the big cow suggested exploring the route that suits you, and I can only talk about the route that suits most people. But before answering, I couldn't help but vomit a person who simply answered "deep learning," "big data," "NLP," and "machine vision." The small direction of each field is very different. Take Natural Language Processing (NLP) as an example. Subdivision has natural language generation, natural language understanding, and language models of different languages. It’s no exaggeration to spend decades on any direction. Just give a few words of answers and what is the difference between buying a lottery ticket...

Therefore, most machine learning practitioners should still be down to earth. Blindly chasing hot spots can easily fall into the trap, and consolidating the foundation, finding the areas that you are good at and the intersection of machine learning can help you become hot in the future job market and become the most talented person in the industry.

What kind of machine learning talents do the industry need in the future? It’s a commonplace to apply the model to people in the professional field, that is, people who cross the field to let the machine learn. Some people will ask if we don't need such a person now? The answer is yes. We need and will need such talents for a long time. There are still various difficulties in the current stage of machine learning. Such a demand will not be a short-lived one. This is a reason for web development. It has also gone through a decade of cycles from fiery to cooling. There is a specific cycle in the development of a field. The threshold of machine learning is higher than that of web development and it is in the sunny period. Therefore, the machine learning experts who are committed to becoming "specialized areas" will not be outdated.

What is a machine learning expert in a specific field? For example, when I used to answer “whether artificial intelligence will replace financial workers”, I mentioned that I studied how to use machine learning to automate part of the audit work in a company, but the biggest difficulty I encountered was my own understanding of auditing. Limited, and other auditors are not very supportive of my work leading to slow progress. So if you have enough machine learning knowledge and a good understanding of a particular field, you can definitely stand on the edge of the advantage in the workplace supply and demand. Take my other answer as an example: "What are the machine learning models used in Fraud DetecTIon?", domain-specific knowledge helps us better explain the results of machine learning models, get bosses and The customer's approval, this is the algorithm has fallen. There are thousands of people who can write code and build models, but there are very few people who understand what they are doing and provide business value from their domain knowledge. So ridiculously, which direction of machine learning talent is the most scarce? A: Every field needs specialized machine learning talents. Your understanding of specific areas is your weapon.

Of course, it is not kind to give the soup without the spoon, so I will give some specific suggestions. Once again, my suggestion is only for friends who are looking for employment. I have different suggestions for taking the research route. I will not repeat them in this article.

After all, machine learning still requires a certain amount of professional knowledge, which can be done through school or self-study. But is it necessary to be fluent in mathematics and good at optimization? My opinion is not needed. The big premise is that you need to understand the basic mathematical statistics. More discussion can look at my answer. "Assam: How to treat" Machine learning does not require mathematics. Many algorithms are packaged. What about a package? "." At the lowest level, I recommend mastering five small directions, which is sufficient for the industry now and in the next few years. Once again, my view of the algorithm is that most people don't build wheels, don't build wheels, don't build wheels! Just understand what you are doing, know what model to choose, and just call the API and the ready-made toolkit.

Regression model (Regression). In fact, the school's curriculum is more about classification, but in fact, the return is the most common model of the industry. For example, product pricing or forecasting product sales require a regression model. The more popular regression method at this stage is xgboost with a number model. The prediction effect is good and the importance of variables can be automatically sorted. Traditional linear regression (mono and multivariate) will continue to be popular because of its good interpretability and low computational cost. How to master the regression model? It is recommended to read Chapters 2-7 of IntroductoryTIon to StaTIsTIcal Learning and take a look at the xgboost package introduction in R.

Classification model. This is a cliché, but you should have a deep understanding of the models that are popular and will continue to be popular. For example, Random Forests and Support Vector Machines (SVMs) are also algorithms that are now commonly used in industry. Perhaps many people can't think of it, Logistic Regression, the classic old algorithm that is common in every textbook in the streets, still occupies more than half of the industry. This section is recommended to look at Li Hang's "Statistical Learning Algorithm" and pick the corresponding chapters.

Neural Networks. I didn't attribute the neural network to the classification algorithm or because it is too hot now, it is necessary to learn to understand. As hardware capabilities continue to grow and data sets become more abundant, neural networks will certainly play in the role of SMEs. This may happen within three to five years. But some people will ask, the neural network contains so rich content, such as structure, such as regularization, such as weight initialization techniques and activation function selection, what level should we learn? My advice is to grasp the classics and master the basic three sets of networks: a. Ordinary ANN b. CNN processing images c. Handling text and speech RNN (LSTM). For each basic network, you only need to understand the classic processing methods. For details, please refer to Chapters 6-10 of Deep Learning and Wu Enda's Deep Learning Network (which has been launched in NetEase Cloud).

Data Compression & Visualization. It is common in the industry to visualize data first. For example, the manifold learning that has been very popular in the past two years has a lot to do with visualization. The industry believes that visualizing is a sharpening of the wood, and compressing high-dimensional data into 2D or 3D can quickly see some interesting things, which may save a lot of time. Learning visualizations can use off-the-shelf tools such as Qlik Sense and Tableau, as well as Python's Sklearn and Matplotlib.

Unsupervised & Semi-supervised Learning. Another feature of the industry is the large amount of data missing, and most of the cases are not labeled. Take the most common anti-fraud as an example, there is very little data on tags. So we generally need to use a lot of unsupervised, or semi-supervised learning to learn with limited tags. To put it another way, the use of reinforcement learning in most enterprises is basically equal to 0. It is estimated that there may not be a particularly wide application in the future for a long time.

The significance of basic skills is that when you face specific problems, you know exactly what weapons you can use. And many of the tools described above have a history of decades and are still alive. So with a 3-5 year span, these tools will still be very useful, and even deep learning algorithms like CNN and LSTM continue to evolve. Whether you are still at school or have started working, mastering these basic skills can be done in a few months to a year or two through self-study.

With basic skills, you can only show that you can output. How can you make your basic skills not the art of killing dragons? Must be combined with domain knowledge, which is why I have been persuading many friends not to blindly switch to machine learning from scratch. And student friends can pay more attention to the areas of interest and think about how machine learning can be applied to this field.

Friends who already have work and research experience should try to use their work experience. For example, don't be the one who is best at investing in machine learning, and the expert who is best at machine learning in the financial field. This is your value proposition. The most important thing is that the basic skills of machine learning are not as high as everyone thinks. There is no need to give up their full-time career change, and the cost of sinking is too high. Through cross-cutting, you can completely save the country by curve, and the disadvantages are advantages. You may have greater industry value than those who only understand machine learning.

To give a few examples around me, one of my friends is doing traditional software engineering research. The year before, he discussed with me how to use machine learning to identify bugs on the history of commits on GitHub. This is a good combination of domain knowledge. If you are financially a person, you can use machine learning to cross the field of your own skills and do strategic research. I have heard countless “claims” using machine learning to implement transactions. Strategy case. Although not trustworthy, a deep understanding of a particular field is often the last layer of paper that breaks through the window. It only understands the model but does not understand the meaning behind the data and data. As a result, many machine learning models only stay in good-looking and not practical. stage.

Looking at it from another angle, people in different fields have an understanding of machine learning that can better promote this technology and break the bubble rumors. For everyone, you don't have to worry about losing your job, you can find your own point of view to find a "golden rice bowl" in this age of deep learning. Therefore, I suggest that practitioners from all walks of life do not have to blindly transfer computer or machine learning. Instead, they should deepen their understanding of the profession and self-study to supplement the basic skills mentioned above, and become a machine learning expert in this field.

Nothing will change, and the technology iteration of this era is very fast. From the beginning of deep learning to the present, but only a short ten years, so no one knows what the next fire will be? Taking deep learning as an example, the two years of very aggressive confrontation generation networks (GAN), multi-lable learning, and transfer learning are still developing rapidly. The theoretical conjecture article on why deep learning has good generalization ability has been recorded in the latest NIPS. This shows that no industry can rely on the old and the old, we still need to catch up with new hot spots. But the scope and field of machine learning is really wide. All of the above are still supervised deep learning. Unsupervised neural networks and deep reinforcement learning are also hot research fields. Therefore, my suggestion is to pay attention to and learn about new hot spots that are mature and have examples.

If you have these basic skills and good field integration, three years and five years is by no means a bottleneck in the profession, and even ten years is still too early. Although the era of science and technology has given us a lot of pressure for change, it has also brought us unlimited possibilities. The technology association is outdated, and the hotspots will always pass, but what will not pass is our enthusiasm for constantly pursuing new technologies and challenges to ourselves.

Stainless Steel Needle Valve Hydraulic Brass

Stainless steel needle valve hydraulic brass is an important part of instrument measurement pipeline system. It mainly includes stop valve and ball valve. Its function is to open or cut off pipeline passage.
Ferrule type instrument valve has the advantages of convenient installation and disassembly, tight connection, conducive to fire prevention, explosion-proof, high pressure resistance and good sealing performance. It is an advanced connection valve in power station, oil refining, chemical plant and instrument measurement pipeline. This type of valve shall be installed horizontally in the pipeline.
Stainless steel needle valve has good sealing performance and long service life. Even if the sealing surface is damaged, it only needs to replace vulnerable parts to continue to use.
During installation, the flow direction of medium shall be consistent with the arrow on the valve body.

Stainless Steel Needle Valve,Solenoid Needle Valve,Motorized Needle Valve,Pnumatic Needle Valve

Taizhou Jiabo Instrument Technology Co., Ltd. , https://www.taizhoujbcbyq.com