AI Research Experience Summary

There is a certain threshold in the field of artificial intelligence research. For beginners, the usual practice is to buy some popular books directly, for example, "Watermelon book", "Flower book", "XX days from the introduction to proficient", "xx days from the beginning to give up" and so on, but most of the books are the basic knowledge, slightly boring and boring, in addition to the content is too abstruse, Beginners may look at it for some time and want to give up. This article in the identity of the past 2 years of their own research experience to share unreservedly to everyone, hope to begin to engage in artificial intelligence research friends help. Start

Find someone you can easily ask questions about.
When you first enter the company, you often hesitate to make some basic questions that may reveal a lack of expertise. But after a few months, my question slowly felt natural, and the questions were carefully drafted. Before that, I will accumulate a lot of problems, but now whenever I have a problem, I will ask questions immediately, so as not to create a backlog of problems and become more and more confused.

Looking for different areas of research inspiration
This is not the time to be alone, pay attention to cooperation. Knowledge is no exception, multidisciplinary communication. For everyone, deciding what direction to pursue might be the hardest part of the study, and here are some of the strategies that I've seen used by researchers with long-term records: interacting with researchers in different fields. Ask questions they are interested in, asking if there is a dataset you want to analyze, and what is lacking in existing technologies. The most effective work in machine learning is the collision with biology, chemistry, physics, social sciences or pure mathematics. For example, I was thinking Matthew
Johnson's article in 2016 nips and Justin
Gilmer in the 2017 ICML article, two articles on the analysis of the mouse behavior DataSet and the application of quantum chemistry; write a simple baseline to get a feel for the problem. For example, try to write some calibration code to control the inverted pendulum. When you write baseline code, you encounter a lot of situations, problems, or occasional ideas that can deepen your understanding of the problem. Expand the experimental part of a favorite paper. Read a paper carefully, understand the methods used and the results obtained, and try to find some perfect places. First of all, consider the simplest extension, and then consider whether the method of the paper is reasonable, the experimental results are not perfect place.

Focus on visualization tools and skills

Running a visual script allows us to quickly verify that the code matches the idea. More importantly, good visualization tends to make errors in ideas and code more obvious and explanatory.
For a practical task, it is difficult to come up with the correct way to solve the problem. If you are using an iterative optimization model (such as depth learning), then drawing a loss function would be a good start. In addition, the visualization of the "black box" method of depth learning can partly explain the neural network parameters it learns. For example, when dealing with a graph model, visualize the distribution of one or two dimensional variables, and when it changes, it can infer a lot of information. Visualization is a barometer of technical effectiveness, and each visual analysis results in a certain amount of feedback on the method or code used.

Figuring out the initial motivations of researchers and papers
Interesting phenomena can be found in academia, the researchers published papers at the same conference, using the same technical jargon, but two of people's research motives can be completely reversed. Motivation is divided into the following three motives--"mathematical" motives, "engineering" motives and "cognitive" Motives: "Mathematical" motivation: the basic attributes and limitations of intelligent systems. "Engineering" motivation: How to develop an intelligent system that solves practical problems better than other methods. "Cognitive" motivation: how to simulate natural intelligence like humans or other animals.

Some papers have more than one motivation, and each researcher's motives cannot remain unchanged, which is related to the interest of the engineer. Good papers and researchers will explain their motives at the outset, but some papers are often not very clear, which requires readers to read carefully, in addition, in their own writing should pay attention to this, in case the motive is not obvious and be rejected or retired. In-depth study

Learn to find a paper
The network is flooded with a lot of artificial intelligence papers, most people will first be published on the arxiv, because the platform can be published before the review, and therefore need to learn to distinguish from. In addition, it is recommended that you follow the dynamics of your favorite researcher on social software. In addition, there are also various meetings that deserve attention. The three major conferences were nips, ICML and ICRL. Other notable general meetings include AAAI, IJCAI and Uai. For each branch of discipline, there are more specific meetings. For example, there are CVPR, ECCV and ICCV in the field of computer vision, and there are ACLs, EMNLP and Naacl in the field of robotics, Corl, Icaps, ICRA, Iros and RSS, and conferences related to theoretical work are Aistas, Colt and KDD. There are also a number of periodicals that are noteworthy, Jair and JMLR are among the most prominent in the field of artificial intelligence, but there are also good papers in the journal Nature and Science.
It is also important to look up some early papers, which are often presented as "classic papers" in a reference paper. Another way to find early papers is to start with a senior professor's personal homepage, which is typically hung on top. In addition, through some search assistants, such as Google Academic, Baidu academic and other query keywords.

How long does it take to read a paper?
about how to read a paper, people often give two suggestions. The first is to read all the relevant papers in the first semester or in the first year of graduate school; the second is to read a large number of papers, do not go to extensive reading, but to find a breakthrough, to come up with innovative ways.
I personally agree with the first proposal, but I do not agree with the second proposal. As long as there is enough time for original research, you should read as many papers as you can. For a professional researcher, it is impossible to always rely on personal luck to discover innovative solutions, and sometimes the methods you think of may have been tried by others, but you don't know it. The vast majority of the researchers are patiently tracking the progress of the research and developing trends, and methodically thinking and solving problems. Reading related papers is also a good way to figure out what the current stage is and what you need to do next.
There is an important hint about reading as much as possible: it is equally important to take time to digest a paper and read a paper, and take notes when reading, rather than to swallow, just quantity, not quality.

Dialogues >> Videos >> papers >> talks
The paper is undoubtedly the easiest source to understand the unfamiliar research theory, but what is the most effective path? Different people may not feel the same way. For me, I find that dialogue (with those who already understand it) is by far the quickest and most effective way to understand. If you can't find a chance to talk to someone like that, you can find video about the problem, such as the author's interview video, which can provide a good view. In addition, when speakers speak to a live audience, they tend to prioritize clarity rather than simplicity. In most thesis writing, the author swaps the priority order, in which the number of words is king, and the background knowledge is too much to be explained by the author. The last is the meeting, the simple talks tend to appear more formal, and the moderator's conversation content may be very valuable.

Beware of hype
Artificial intelligence has achieved a series of results that have attracted public attention, and more people have been involved in this field, which in turn has led to more breakthroughs in artificial intelligence. The whole cycle is benign, but one side effect is that there is a lot of hype. Views to get clicks of reporters, hot money investors, entrepreneurial companies are exaggerated speculation bubble culprit. Therefore, when we read news or papers, we should pay attention to the "title Party" to avoid being misled.
In the 2017 question and answer session of Nips, a famous professor with a microphone (on behalf of the police hype) cautioned the author to use the word "imagination" carefully in the title of the paper. This is the same as we read the news, the title is very attractive, but the contents of the content is not related to the title, so that readers disappointed. Reading paper is also so, to prevent speculation, we need to do is based on the experimental methods and results to assess whether a paper to help themselves. Research is a marathon.

Always making progress.
In the early exploration of research projects, I usually spend a few hours brainstorming, hoping that some vague direct can guide a specific direction. Sometimes there is no progress in the project, but groping in the dark is part of the whole research process. When you don't know what to do next, you can write down the most ambiguous thoughts based on what's already in place and make one by one out of the writing process (write out the reasons for the exclusion). In the absence of any idea, you can take the form of reading or communicating with colleagues to get inspiration.

Learn to discriminate and stop from a dead end
Daniel usually spends more time on good ideas, and the ability to distinguish between good and bad ideas depends to a great extent on personal experience. Nevertheless, researchers at any level are constantly confronted with the following decisions: whether the research ideas are flawed, whether they should be saved or further supported, and whether the ideas presented are completely abandoned. Especially in the early days, researchers stuck to a dead end for a long time rather than giving up. While giving up means that the time spent in the past is in vain, sometimes you have to know how to stop in time. Learn what

Writing
Some of the early career advice given by Daniel is: writing. You can write blogs and essays at ordinary times, but it is more important to record your thoughts. Because writing helps us to understand and think about the relevant knowledge.

Mental health and physical health are prerequisites for scientific research
It is not a good habit for academic researchers to encounter problems such as staying up late and eating in the process of pursuing scientific discoveries. Many PhDs begin to get bald, and even masters begin to lose their hair. Exercising and emptying yourself is an investment in research, not a hindrance to research. Working 4 hours after 8 hours of sleep is much more efficient than 8 hours of work for 4 hours. Sometimes you will encounter jam, even if you can not make a trace of progress, this time the proposal to leave the job, a little activity and do a long breath, vent their own.