A machine learning doctor's advice [go]

Purely reproduced, there is reference value, but also to encourage!
Know the original

The main question does not seem to be clear that he is a PhD student.

1. First of all, take an outsider's intellectual perspective and classify your mentor as follows:
① Inner circle Daniel, high-level paper many, at present oneself still in the pro-active for participation in scientific research work, please turn 2
② Resume There are some high-level papers, but it does not seem to be Daniel, please turn 3
③ other conditions, please turn 4

2, this kind of situation is more ideal, personally think you about the paper doubts, in fact, as much as possible with your mentor to communicate. Not to mention such a big topic of scientific research, just to write a paper, in fact, there are a lot of skills and unspoken rules, with their own to try the wrong, better from the mentor to learn more.

3, this situation, the need to take more of their hearts, the instructor may be busy with administrative reasons such as not to direct your guidance, then you must not let go. The reality of the domestic situation, the tutor is too many students, the vast majority of things are not likely to help you consider. Specific to the study of machine learning, you can discuss with the instructor to determine a general direction, and then find yourself a specific problem (if the instructor has not been able to help you to confirm the general direction, please turn 4). In formulating research questions, these questions must be answered with certainty:
① you can get the experimental source data used in the most cutting-edge papers (refer to the latest conference papers, machine learning field, please refer to ICML NIPS aaai CVPR, etc.)
② for these experimental source data, whether your computing resources can be effectively hosted (for example, a machine that was configured only five years ago, the depth learning direction will be doubly difficult to get into)
③ to the question itself, and the abstract idea behind the question, do you personally agree and love (when you study a type of algorithm, if you don't like it, the speed of ideas will become noticeably slower)
④ The research question, whether there is a continuous emergence of nearly 3 years of conferences and nearly 5 years of periodical articles
The above questions are a necessary (but not sufficient) condition for the effective development of a research project. As you can see, the answer to these questions, like the ones that have been answered, requires that you have an overall grasp of the research that is relevant to this small problem, that it is a good way to write a survey yourself, and that the reference to more than 100 of the literature is probably a primer. At that time, you will have a variety of ideas to wait for the realization, of course, there are about 40% of these ideas are not reliable, there are 40% people have done, there are 10% you do not have the ability to do, will be converted into your paper is the remaining 10%, but these are relatively simple.

4, "First-class university students, as well as mentors enough of the students, you do not have to look down, meaning not very much"
If you find yourself in such a situation, you must be vigilant. This situation is extremely detrimental to doctoral students, and a slight carelessness will make themselves in the next few years extremely painful (personal experience). So, I have the following suggestions:
① pay attention to administrative time. Even if you are a hard-working and self-disciplined person, you must also pay attention to whether your daily research time is used in matters related to your own research. A few simple examples:
A) A large amount of time for the horizontal project of the instructor, the horizontal project is not related to the research topic or the engineering nature is too strong.
b) Work related to data acquisition/data preprocessing is performed at the request of the instructor, and no cutting-edge technology is used.
c) The tutor does not care, but is not allowed to look for research directions by himself.
② more hands-on, less study. What we mean by "less learning" here is that you don't need a Python programming guide for three weeks because you're using a toolkit that needs to use a Python language you've never been in touch with. This "learning" is basically a waste of time, about Matlab, Python, Hadoop, OpenCV, NLTK ... These familiar tools, please use official documents, Baidu and Google, according to your needs, while doing learning.
③ and mentor complain. As the saying goes to cry the child has milk to eat, do not feel very strong and very lofty on the head of their own dry, your mentor may not be able to research, but you have to recognize the reality: only he can help you. Even if it's just a good computer or a bigger table, it's all about your research.
④ write more essays. Especially in the early days of reading Bo, do not worry about the level of the paper is not afraid to move handwriting, the correct approach is that after the completion of the survey process, should maintain a continuous paper writing, as long as you have not written paper for more than 10 days, you should give yourself a red light. The reasons are as follows:
A) The paper is your graduation chips, no chips, you have nothing.
b) The paper as long as not too bad, write to always be able to send, may be retired several times, perhaps cast a low-grade periodicals, but can be issued.
c) Writing a paper is a need to practice, before writing a good paper, generally need to write a lot of bad papers.
D) Unless your talent is different or English is very good, no matter how good your innovation, in terms of language, you have just begun to write an English paper in the foreign reviewers are basically Dog.
e) be reject several times, know how to write the paper.
⑤ about the main topic, feel that various algorithms have been improved, do not know where to start the problem. In fact, mainly due to the problem of research and related methods of unfamiliar, do not need to worry about the hole has been filled with special problems. You can take a peek at this paper: Manuel Fern andez-delgado et.al. Do we need hundreds of Classi ers to Solve Real world Classi cation problems? , Journal of machine Learning 15 (2014) 3133-3181. There are 179 kinds of algorithms used for classification, these algorithms are also made of papers, it can be said that most of them in fact there is no very good application value. So do not be too nervous, any little bit of improvement, ideas in a reasonable packaging can be a high-quality paper. For the master, you can think of some ideas combining, such as someone using Method 1 to solve problem A, some people use method 2 to solve problem B, then I use Method 2 to improve the method 1 to better solve problem A, this is the point of the paper.
⑥ 工欲善其事 its prerequisite. From the paper review and download, document management, note management, data collection and collation, experimental tools, paper writing process and other aspects, more optimization of their own work flow, save time even to sleep is good. Repeat the above mentioned, please use Baidu and Google, of course, if you have a reliable brother and sister can also.
⑦ Learn more. This is not in conflict with the above ②, the study here refers to the basic knowledge system that is closely related to your scientific research ability, rather than how to realize your simple idea in Python. Whether it's a classic textbook (PRML, Mlapp, ESL, etc.) or a classic Open class (Ng machine learning, Heights machine learning Cornerstone + technique), it takes time to learn. Otherwise you will find that you think you understand the paper, in fact, you still do not understand, so you can not make your own research.

A machine-learned doctor's advice [go]