1. 1. To master the necessary mathematical foundation, there are actually three courses in college: A. Calculus B. Linear Algebra C. Probability and Statistics. If you have spare capacity, you can also look at things that optimize the direction, but this is not mandatory.

1.2. Master some core courses of computer basics, such as database, parallel operation and discrete mathematics (temporarily put in the computer column). As for programming skills, I don't think it is necessary to be particularly powerful. If the ability is limited, you can lower the level of courses such as operating system and computer structure. Learn to grasp the big and let go of the small in life, and don't force yourself to master everything.

1.3. Learn Python well, and learn programming habits (PEP8) and grammar sugar on Python. At the same time, you can learn more about Python's corresponding data science/machine learning tool libraries, such as pandas, numpy, scipy, sklearn and so on. Even if you don't do machine learning in the future, the knowledge of the glue language Python can still be of great help. Another reason for choosing Python is that most deep learning frameworks, such as TensorFlow/Theano/Keras/Pytorch, are based on or have Python interfaces.

1.4. Start learning basic machine learning. The recommended methods are: a. Watch Andrew Ng's machine learning course on Coursera. B. Start reading basic books on machine learning (such as Collective Programming Wisdom, Python Machine Learning and Introduction to Statistical Learning). At this stage, the most important thing is not to bite off more than one can chew. If you browse Zhihu, you will find that everyone says that you must read Elements of Statistical Learning, PRML and so on. I admit that reading such a book will be helpful, but it is not suitable for reading directly at first, because it may make you "give up when you get started." With some basic knowledge, you already know what you need to do next. I hope to give you the right to choose, not recommend a bunch of courses and books. Of course, if you want to go deeper, you can read Zhou Zhihua's Machine Learning and Li Hang's Basic Statistical Learning in Chinese, and you can also read Essentials of Statistical Learning and Deep Learning in English. The focus of this stage is to form a systematic knowledge context, remember to bite off more than one can chew, remember!

1.5. To learn English well, you should at least lay a good foundation in reading and listening. Although China has done a good job in the field of artificial intelligence, the mainstream books, periodicals and conferences are all in English. We can accept the translated version, but the best way is to be able to read it directly. Even if you don't do machine learning in the future, English reading ability will still be of great help.

2. Practical experience:

2. 1. Try to get in touch with scientific research and enter the laboratory as soon as possible. Generally speaking, you should have the basic knowledge of machine learning in your junior year, although it is still very shallow. At this time, you can recommend yourself to the teacher/senior/senior to enter the laboratory, even if you work for free and do basic coolies. Entering the laboratory has two obvious advantages: a. You will have a deeper understanding of a small direction. Generally, undergraduates don't need to do pure theory in the laboratory, but need to compare small directions such as machine vision or natural language processing (NLP), so this is a good opportunity to learn more about one direction. B. You can know whether you are suitable for this field by supplementing your research experience. If you are lucky, you may even be one of the authors of the paper, or even go to a meeting (meet the industry leaders when traveling at public expense). This is very helpful for further study and further study abroad. Having scientific research experience and papers is a great bargaining chip, which is absolutely beneficial and harmless for finding a job.

2.2. If you are interested in scientific research, you can try it as soon as possible. Most of the methods we read in books are actually more or less ideal models, and even many of them are out of date. For example, in most textbooks, the activation function of neural network is still sigmoid, and the industry has long since stopped using sigmoid. The biggest threshold for machine learning beginners is that they have learned a lot of knowledge, but they have no chance to use and test it. Internship as early as possible can give you a more intuitive feeling and prevent them from having only one skill to kill dragons.

2.3. Maybe scientific research and internship opportunities are sometimes hard to get. At this time, you must find your own project to do, and use interest to drive the project. Better methods include taking part in Kaggle competition, Tianchi competition, or applying machine learning to things you are interested in. I once saw Zhihu use machine learning to judge whether the last forty chapters of A Dream of Red Mansions were written by Cao Xueqin, not to mention whether the articles were rigorous or not, but this is a good example of using interest to promote practice.

3. Social sciences and humanities:

3. 1. As a rapidly changing field, machine learning should have its own "persistence" and "taste". For a simple example, the glory of deep learning now needs to be attributed to the persistence of a group of scientists in the trough of neural networks. But at the same time, even if you only talk about machine learning, don't believe that only deep learning is the best, and you can't stick to your own opinions. If you are interested in machine learning, don't think that today's network security is good, and think that human-computer interaction (HCI) is the most promising tomorrow. Chasing hot spots often produces bubbles.

3.2. Read more books in different fields, such as social science, economics and humanities. Because the starting point of data science is based on data, the end point is to extract opinions and provide feedback. The viewpoint does not show the height at a glance like a number, but needs empirical multi-angle analysis. Many people think that it is enough for computer majors to read mathematics and papers, but in fact science is always intertwined with society. For example, whether artificial intelligence should be applied to the military is a hot issue in Zhihu recently, but to answer this question, you need to have enough non-computer knowledge reserves. There is no good or evil in science and technology, but people do.

3.3. Lower your figure and communicate with others more. In fact, this is a bit far-fetched, but I found that the small partners who do machine learning are generally very individual, and many of them are introverted (including myself), which may be applicable to most friends in science and engineering. Even so, I suggest you let go a little bit, because most data-oriented jobs need interaction, such as data analysts and data scientists.

4. Write at the end:

With the rapid development of science and technology, it is good to chase hot spots. But as I just mentioned, in this impetuous era, no matter what direction you choose, the most important thing is the ability to think independently and the courage to discard the false and retain the true. If a thing cannot be proved or falsified, it must be in doubt. Universities are both the best and the key period to cultivate good scientific literacy.

So seeing so many people sharing their experiences, what I hope most is that you don't rush to accept it all, and don't deny it all because you don't like it. Slow down and think about it. This is probably the correct attitude to do scientific work.