Author: Artificial Intelligence Laboratory edit all graduate students: David Chapman Version: 1.3 Date: September 1988 Translator: Liu Quanbo thrust of Beijing Normal University, School of Information 2000 PhD Abstract is to explain how to do research. We provide these recommendations that are extremely valuable to do research itself (reading, writing, and programming), understanding the research process, and the beginning of love research (methodology, topic, selection, selection, and emotional factors). Copyright 1987, 1988 Author Copyright Note: Working Papers of Artificial Intelligence Lab for internal communication, including information that is too preliminary or too detailed to publish. Unlike formal papers, all references will be listed. 1. Introduction No God Dan Miao can guarantee success in research, this article only lists some informal opinions that may be helpful. Who is the target reader? This document is mainly written for graduate students in the MIT Artificial Intelligence Lab, but it is also valuable for artificial intelligent researchers of other institutions. Even the researchers in the artificial intelligence, you can also find some part of yourself. how to use? It is too long to read this article, it is too long, it is best to use the way. Many people think that the following methods are very effective: I will read it quickly, then select some of them to study with their current research projects. This document is roughly divided into two parts. The first part involves various skills required to researchers: reading, writing, and programming, and so on. The second part discusses the research process itself: what is going on, how to do research, how to select questions and selection, how to consider emotional factors in the study. Many readers reflect, from the long run, the second part is more valuable than the first part, and more interest. How to lay the foundation of AI research. The important AI journals are listed, and some readings are given to becoming a member of the AI research: keep in touch with relevant personnel, they can keep you keeping track of the cutting edge, know what material should be read. Learn knowledge in the relevant fields of AI. There are basic understandings for several areas, and it is necessary for one or two fields. How to do research notes. How to write journal papers and graduation thesis. How to write reviews for drafts, how to use other people's reviews. How to publish a papers. How to do research reports. It is related to the program. The AI program design is different from the process of usual habits. How to choose a tutor for the most important issues of the research workfare. Different mentors have different styles, this section will help you find a suitable mentor. The tutor is you must understand how to use resources. About the graduation thesis. Graduation Thesis will occupy most of the graduate career, this part involves how to select questions, and how to avoid waste time. The research methodology has not been completed. Perhaps the most important section: Impressive factors in research, including how to face failure, how to set goals, how to avoid unseason, keep confidence, enjoy happiness. 2. Read the time reading of many researchers. You can learn a lot quickly from others' work. This section discusses the reading in AI, which will discuss other subjects related reading in the fourth section. Read the literature and start on today. Once you start writing a papers, there is not much time, and the reading is mainly concentrated in the literature related to the theme of the paper. In the first two years of graduate students, most of the time should be used to do course work and play foundations. At this point, you can read the textbook and published journal articles. (In the future, you will mainly read the draft of the article, see the section 3). The reading amount required to make a solid foundation in the field is a persistent.
But since AI is just a small research area, you can still spend a few years to read the most essential part of many of the quantities published in the art. A useful little skill is first to find out the most essential papers. You can refer to some useful bibliographies: such as graduate courses, other schools (mainly the Stanford University) Postgraduate admission procedures, these can make you have some initial impression. If you are interested in a sub-domain of AI, the senior grade graduate students in this field will ask for the most important ten papers in the field, if it can, borrowed copy. Recently, there have been a lot of careful editing papers, especially Morgan-Kauffman published. The AI laboratory has three internal publications series: Working Papers, Memos, and Technical Reports, formal degrees of extent increase, can be found on the eight-layer shelf. Review the publication in recent years, copying those very interesting. This is not only because many of them are very significant papers, but also important for the progress of the work of laboratory members. There are a lot of journals about AI, fortunately, only part of it is worth seeing. The core journal is Artificial Intelligence, and there is also written "The Journal of Artificial Intelligence" or "Aij". The papers that truly value in the AI field will eventually be cast to Aij, so it is worth browsing the Aij of each period; but there are many papers that have a lot of the paper. Computational Intelligence is another journal worth watching. Cognitive Science also publishes a lot of significant AI papers. Machine Learning is the most important resource in the field of machine learning. IEEE PAMI (Pattern Analysis and Machine Intelligence) is the best journal related to visual journals, each period has two or three valuable papers. International Journal of Computer Vision (IJCV) is the latest founder and is currently valuable. Robotics Research's articles are mainly about kinetic, sometimes there is an epoch-making intelligent robot papers. IEEE ROBOTICS and Automation occasionally have a good article. Every year, you should go to the school's computer science library (at one level of MIT's Tech Square), please read the AI technology report published in other institutions and select the careful reading of yourself. Reading the paper is a skill that needs to practice. It is impossible to read all the papers completely. Reading papers can be divided into three phases: The first stage is to see if there is something interested in the paper. The AI paper contains a summary, which may have content, but it may not be or concluded, so you need to jump, this look at it, know what the author has done. The Table Of Contents, Conclusion and Introduction are three focus. If these methods can't, you have to browse quickly. Once you understand the probably and innovation points of the papers, you can decide whether it is necessary to carry out the second phase. In the second phase, we must find out the part of the paper truly content. A lot of 15 pages can be rewritten as a page of about a page; therefore need to find those places that are really exciting, which is often hidden in a place. The author is not necessarily what you are interested in, and vice versa. Finally, if this paper does value, return to the reading.
When reading the paper, keep in mind a question. "How should I use the paper?" "Is it true like the author claim?" "If ... What happens?". What conclusions in understanding the paper are not equivalent to understanding the paper. Understand the papers, it is necessary to understand the purpose of the papers (many of them are implicit), whether assumptions and formalization are feasible, the papers pointed out, what problems are there in the fields involved in the papers, the authors of the authors What is the difficult point mode of continuous appearance, what is the point of view expressed in the paper, and the like. It is very helpful to link reading and programming. If you are interested in a field, after reading some papers, try the "toy" version of the procedure described in the paper. This will undoubtedly deepen understanding. Sadly, many AI laboratory is naturally lonely, and members mainly read and quote their work in their school laboratories. It is important to know that other institutions have different ways of thinking, it is worth reading, treating seriously, and quoting their work, even if you think you know what they are in your mistake. Some people will always give you a book or a paper and tell you that you should read because there is a very flashing place and / or can be applied to your research. But waiting for you to read, you find that there is nothing special flash, just barely available. So, confusion is coming, "Why don't I?" I missed what? ". In fact, this is because your friend is reading some ideas that have already formed in the minds in reading books or thesis, and it is seen that there is a value for your research topics. 3. After two years of establishment, there have been some ideas for the subsequent areas of themselves. At this point - or earlier - joining Secret Paper Passing Network is important. This informal organization is the reflection of artificial intelligence truly. The work of booting the trend will eventually become a formal published papers, but at least after the cattle understands that it is, that is, the cow is at least one year for new ideas. How does the cattle find new ideas? It may be listened to a meeting, but most likely come from Secret Paper Passing Network. Below is a general situation of this network work. Jo Cool has a good idea. She will still be integrated with other work, written a draft paper. She wants to know how this idea is, so she sends the copy of the paper to ten friends and ask them to comment. Friends think this idea is great, but also points out the mistakes, then these friends have copied the papers to their respective friends, so continued. After a few months, JO has made a lot of revisions and sent to AAAI. After six months, the paper was officially published in five pages (this is the space allowed by the AAAI Conference). Finally, Jo started to organize the relevant program and wrote a longer papers (based on the feedback from AAAI published papers). Then send it to the AI journal. The AI journal will spend about two years, review the paper review, including the author, and the time spent on the paper, and the corresponding publishing delay. Therefore, in the ideal case, JO's idea finally published in journals for approximately three years. So the cattle is rarely learned from the journal articles published in this field, it is too late. You can also become a cattle. Below is some of the 窍 建 建 学 学:: There are a lot of email lists that discuss a AI sub-domain (such as connectivity or vision), select the list of interest to join.
When discussing your own thoughts with people who are familiar with those in the art, they are likely to evaluate your thoughts directly, but said: "Do you have read a certain?" This is not an asked, but I suggest you go to read. A certain literature, it is likely to have a relationship with your thoughts. If you haven't read this document yet, you will get the details of the document from the master talking to you, or you will borrow a copy from him. When you read a papers that makes you feel excited, copy five other five people who are interested in. They may feedback back very good suggestions. This laboratory has many discussion groups for informal (sustainable development) papers for different subsequent fields, and they will discuss all the papers read by everyone every week or every two weeks. Some people don't mind others to go to the desk, that is, go to read the papers who have to read or often read them on the desk. You can go over and have you interested. Of course, you must first get the owner's license, you have to know that some people really disgust what others turn their own. Try those who are close to the people. Similarly, some people don't mind if you look down on their files. In the laboratory, there are many people who have learned a deep learning, and there are many babies in their file cabinets. This is usually a way to find papers faster than using the school library. As long as you write something, you will send a copy of the draft to those who may be interested. (This also has a potential question: Although there is very little plagiarism in the AI field, you can do it on the first page "Please don't photocopy or quote" to make some precautions.) Most people will not Read most of the papers received, so if you only return to you, you don't care too much. You can repeated several times - this is necessary for journal panels. Note that except for your mentor, it is generally rarely to the same manuscript. When you write a papers, give the copy of the paper to those who may be interested. Don't think that people will naturally go to read the journals or conferencing of the published papers. If it is an internal publication (memo and technical report), it is not easy to read. The more people you keep in touch, the better the effect. Try to exchange papers with people in different research groups, AI laboratory, different academic fields. Make yourself a bridge that has not contacted the two research groups, so that it is very fast, you will take a big papers on your desk. If a paper quotes some things you are interested in, do your notes. Maintain a log of your own reference to the reference. Go to the library to see if you can find these papers. If you want to know the development trajectory of a topic, you can be intentionally to do a referenced "reference" map. The so-called reference diagram is a network of guidelines: Thesis A references B and C, B reference C and D, C reference D, and so on. Note those papers that are often cited, which is usually worth reading. The reference chart has a wonderful nature. One is often the study group that studies the same topic does not understand each other. You searched this picture, suddenly found that another part of the way, usually appeared in different schools or different methods. It is very valuable as much as possible to understand many ways, which is better than a very deep understanding of some way. Temporary shelves. Talk to others. Tell them what you are doing, and ask what people do. (If you are shy to discuss your own ideas with other students, you must insist on talking, even if there is no idea, discussing themselves thinking to do excellent papers. This will naturally guide what to do next. ) There is an informal lunch seminar every day at noon. In our laboratory, people are accustomed to work at night, so they can be discussed with the loose group of people in lunch. If you communicate with the outside world - do demonstration or participate in the meeting - to print the business card, mainly to make your name easily.
From a certain time, you will start to participate in the academic conference. If you really participate, you will find a fact that almost all conference papers are more gathered or stupid. (This reason is very interesting, but it has nothing to do with this article, not discussion). Do you still go to the meeting? Mainly in order to get people outside the laboratory. The people outside will spread news about your work, invite you to report, tell you the characteristics of the academic atmosphere and researchers in a place, introduce you to others, help you find a summer work, like this. How to get to someone else? If you feel that someone's thesis is valuable, run up, say: "I appreciate your paper" and ask questions. Get the opportunity to work in summer work in another laboratory. This way you will meet another group, perhaps you will learn another way to look at things. You can ask how the high grade classmates get this opportunity, they may have worked in the place you want to go, can help you contact. Usually, you can only do things in the AI field, don't know about things outside the AI field, as if some people still think so. However, now requires good researchers to understand a few related areas. The calculated feasibility itself does not have sufficient constraints for intelligence, other areas given other forms of constraints, such as psychological acquisition experience data. More importantly, other research fields give you a new tool for thinking, seeing a new way of intelligence. Another reason for learning other fields is that AI itself does not evaluate the standard of research value, which is all borrowed from other fields. Mathematics will be made as progress; the project will ask if an object is reliable; the psychology requires repeated test; philosophy has strict thinking; All of these standards are sometimes played in AI, familiar with these standards help you evaluate the work of others, go deep into your work and protect your job. After six years or so, you can get the MIT's PHD. You can lay a solid foundation in one to two non-AI fields, with reading levels in more areas, and must have a certain degree of understanding of most of the content. . Here's how to learn some of the methods you know very well: Elective a graduate course, which is very reliable, but usually not the most effective way. Read the textbook. This method is also good, but the knowledge of textbooks is often outdated, and there is also a high proportion of rhetoric with content. I find out what the best journal in this field is to ask the high people in the field. Then I find out the article worth reading in recent years, and track related references. This is the fastest way to feel the field, but sometimes you may have a mistake. Find the most famous scholars in this field and read the books they have. Bubble with graduate students in this field. Refer to the foreign school studying the course table in this field. If you visit the Institute of the Institute, choose useful literature mathematics may be the most important discipline that needs to be understood. For those who work in vision or robots, it is more important. For the system-centric work, on the surface, it is not relevant, but mathematics will teach you useful thinking. You need to read theorem, if you have the ability to prove theorem, you will have a deep impression in most people in the art. Few people can self-study mathematics, light to be a listener is not enough, but also have to do the topic. Elective as much as possible as much as possible, and the courses in other fields are also easy. Computer science is based on discrete mathematics: algebra, chart, and so on. If you want to work in reasoning, logic is important. Logic is used in Mit, but in Stanford and elsewhere, this is the mainstream method of understanding thinking. So you have to have enough logic knowledge so you can protect your own point of view. It is enough to choose one or two courses in MIT Mathematics.
If you study interest in perception and robots, then not only discrete mathematics, but also need continuous mathematics. In terms of analysis, differential geometry and topology have solid foundation will provide you with the most common skills. Statistics and probability are only generally useful. Cognitive Psychology and AI share almost identical views, but practitioners do have different goals, they are mainly experimentation rather than writing procedures. Everyone needs to know some knowledge of cognitive psychology. In Mit, Molly Potter opened a good primary graduate course about cognitive psychology. If you want to do work about learning, then developing psychology is very important. Develop psychology is also useful in general sense, it can tell you what is easy for human intelligence. It also gives a cognitive model for cognitive architecture. For example, work on children's language acquisition adds a solid constraint on language processing theory. In Mit, Susan Carey has opened a very good primary postgraduate course for developing psychology. Part of the "soft" in psychology, such as psychological analysis and social psychology, the impact of AI seems to be small, but has potential significance. They will give you a very different understanding of people. Social science like sociology and anthropology can play similar roles. It is useful to have a variety of views. You need to learn from the above discipline. Unfortunately, it is difficult to distinguish between these areas which are both garbage. Go to Harvard to learn: For MIT's students, it is easy to cross the courses of Harvard. Neuroscience tells us that the human body can compute hardware. As the rise of neuroscience and linkage can be calculated, AI has a very large impact. MIT's Brain and Behavioral Science Department offers a very good course, the visual (Poggio, Richards, Ullman), Mobile Control (Hollerbach, Bizzi) and Ordinary Neuroscience (9.015, taught by an expert group). If you want to study natural language processing, linguistics is very important. Not only that, but it also contains many constraints about human cognition. In mit, linguistics is mainly responsible by the Chomsky College. You can see if it is in line with your interest. George Lakoff recently published book "Women, Fire, And Dangerous Things" can be used as an example of another research program. Engineering, especially motor engineering, has been used as a research area as a research institution. Our laboratory has added a lot of requirements that need to do something in the cultivation procedure, such as analysis circuitry. Understanding EE also helps to build a custom chip or debug your power on your LISP machine. Physics has a powerful impact on those interested in perception and robots. Philosophy is a frame that is invisible in all AI fields. Many Ai work has a philosophical impact. Learning philosophy can also help you use or read the views used in many AI papers. Philosophy can be decomposed along at least two orthogonal axes. Philosophy is usually a philosophy of something; philosophy of thinking and language is more relevant to AI. There are also a variety of philosophical schools, from a relatively large range, philosophy can be divided into analytical philosophy and mainland philosophy. Analysis of philosophy about thinking and consistent with most researchers in the field of AI. Continental philosophy has a very different way of viewing many things happened often. It has been used by Dreyfus to prove AI is impossible. Not long ago, several researchers believe that the Chinese philosophy and AI are compatible, providing another way to solve problems. MIT's philosophy belongs to the analysis philosophy, and the Philosophy of Science is deeply influenced by Chomsky in linguistics. It seems that there is too much thing to learn, is it? That's it. Be careful: thinking about all X, "Only I know more about X, this problem will become easy". To know, something that needs to be further understood is never finished, but in the end, you still have to sit down and solve the problem.
5. Many scientists have a habit of doing research notes, you should also do this. You may have been told from the fifth grade, and you should remember your notes for each science. Different ways of writing are suitable for different people, they can do online notes, record on notebooks or notebooks. It may be necessary to have one in the laboratory, and there is one at home. Record your own ideas on the notebook. Only you will read it yourself, so you can remember more casual. Record your own thinking, problems encountered in the current work, possible solutions. Summary of references that may be used in the future. Turn on your own notebook regularly. Some people will be a monthly summary to facilitate future references. Things in the records in the notes can often be used as the backbone of a paper. This will make life easier. Instead, you will find a rough paper - title, summary, division header, and body fragment - is an effective way to record your current work, even if you are not ready to turn it into a true papers. (You may change the idea after a while). You may find that Vera Johnson-Steiner's book "Notebooks of the Mind" is useful. This book is not a literature that describes how to do notes. It describes how innovative ideas appear as the accumulation of thought pieces. 6. There are many reasons for writing writing. During the process of the entire study, you need to write one to two (which depends on the provisions of your system) graduation thesis to get PHD or MS. Diligent writing is not only the opportunity to practice. The academic rules are either published or rot. In many fields and schools, this usually begins when you become a professor, but many graduate students in our laboratory have begun to publish their paper. Encourage publishing and distributing papers is a good policy. Write down your own ideas is a good way to adjust your ideas. You will often find self-righteous ideas, once written down, it is meaningless. If your work is to pay for yourself, you must publish it. This is also the basic responsibility of research. If you write wonderful, there will be more people to know your work. Ai but it is difficult to do it, you need to get feedback from others frequently. Comment on your paper is the most important form. Everything is going to do it. Read books on how to write. "Elements of Style" for Strunk and White is how basic should be introduced. Claire's "Houghton Mifflin is about how to edit in the sentence level. Jacques Barzun's "SIMPLE AND DIRECT: A r HETORIC for WRITERS" (HARPER AND ROW, 1985) is related to how to essay. When writing the paper, read the books that write superb, and think about the author's syntax application. You will find unknowingly, you have absorbed the author's style. To become a master, you need to pay a lot. After a few years, you have to endure and seriously treat others. In addition, there is no shortcut to go. Writing sometimes is very painful, it seems that it is distracted from "actual" work. But if you have already mastered writing skills, you will be very fast. And if you regard writing as an art, you can get a lot of fun. You will definitely encounter ideas to block, this has a lot possible reason, and there is no way to avoid it. Pursuit of perfection may cause ideas to block: no matter what to start, I always feel not good enough. To understand writing is a debugging process. Write a draft first and return to the revision. Write a draft contributing to the road, if you can't write a text, then write an outline. Differentially refine until it is easy to write the contents of the child portion.
If the draft can't be written, hide all the windows being written, then enter the things you think in your own head, even if it looks like garbage. When you have written a lot of text, reopen the window and edit the things you just wrote. Another error is to read all the content sequentially. Usually, you should write the core content of the paper, and finally the introduction is. Another reason that caused author thinking to block is unrealistic to think that writing is very easy. Writing is time consuming, if you find yourself can only write a page every day, don't give up. Perfectionism may lead to a papers that have been good enough to be polished. This is a waste of time. (This is also a manifestation that is interested in escaping the study). Watch the paper as a sentence when talking to others in the field. In the conversation, not every sentence is perfect. Few people will look forward to your own conversation is all stories, and is the last exchange with each other. Write a letter is a good practice. Many technical papers, if its style is more similar to the letter to a friend, then there will be great improvements. Adhere to the diary is also a way to practice writing (will also make you experiment with more style, not just technical papers). These two methods have other substantive effects. A common trap is spent for many times to pursue rhetorical instead of content. To avoid this. LaTex is not perfect, but it has a lot of modifications you need. If this is not enough, you can borrow some words usage from other people engaged in this study. Many sites (such as MIT) maintain a library of writing rhetorical. Clear what you have to express. This is the most important factor in clear writing. If you wrote a bad thing, and don't know how to modify it, this is likely because you don't know what you want. Once you know what you want, you will say it. The writing of the paper is beneficial to the reader to find what you do. Whether it is a paragraph organized or artificial organization, it is necessary to put the core part in front. Writing a summary. Make sure the abstract has reflected your good idea. Make sure you understand what your innovation is, then express it in a few words. Too many papers are only generally introduced the papers, saying that there is a good idea, but nothing. Don't use big words to sell your work. Your readers are very good people, integrity and self-esteem. In contrast, don't apologize or reduce your work. Sometimes you realize a clause, sentence or paragraph is not good enough, but I don't know how to modify it. This is because you can't get it in the dead alley. You need to return to rewrite this part. This situation in reality rarely occurs. Ensure that there is a central idea in your own papers. If your program solves the problem X in 10 milliseconds, tell the reader how you do it. Don't just explain how the system is built, what is it, but also explains its working principle and value. Writing is to people, not machines. Therefore, it is easy to understand that the light view is correct. Don't rely on readers yourself, unless it is the most obvious inflation. If you explain the working principle of a little bit on the footnote on page 7, then there is no further interpretation in the twenty-third page, it will be quoted, and if the reader feels confused, it is not worthy of strange. The formal papers should be difficult to write. Do not imitate the literature in mathematics, their standards are as little as possible, so that the readers feel more and better. This does not apply to AI. After writing a paper, delete the first paragraph or a few words. You will find that is a general language that is unrelated to the content, a better introduction statement at the beginning of the first or second paragraph. If you wait until all work, you will lose a lot. Once a scientific research project is started, it is necessary to develop such habits: writing interpretation of the current working progress or informal papers to learn from a month. Start with the records in your research notes. It takes two days to write down - if you spend longer, you are a perfectionist. Share the paper with your friends. Writing a draft - not to be referenced.
Ten the copies of the paper and give those people who are interested in (including your mentor). Compared with the formal papers, doing this with a lot of the same benefits (comment, clarifying ideas, writing exercises, etc.), and in a sense, there is no need. Often, if you do it, these informal papers can be used as the backbone of formal papers, which is a journal article from the AI Laboratory Working Paper. Once you become a member of Secret Paper Passing Network, there will be many people to give you a copy request comment. It is very valuable to get the comment of your own papers. So your comment is more, the more support you, you will also receive more people's comments on your paper. Not only that, learning to evaluate others' papers to help you choose. Comment on the paper is an art. To write a useful comment, you need to read two papers. The first time I understand its idea and start comments second. If someone has repeatedly made a mistake in the paper, don't mark each time. But what is going to figure out what it is. Why do he do this, what can you do, then clearly indicate or communicate privately in the first page. The author of the paper will follow the principles of minimal modification when merging your comments. If you can, only one word is modified, and you can modify a phrase and then modify the entire sentence. If some of his papers makes him have to modify the entire paragraph, the whole section or even the whole papers, to point out the letters of the big font, so that he will not ignore. Don't write devastating criticisms such as "garbage". This is no help for the author. Take time proposes a constructive advice. It is necessary to think of the author. Comments have many kinds. There is a comment on expression, there is a comment on the content. The reviews of expressions can also be very different, can be the school's punishment, punctuation, spelling error, word loss, etc. You should learn some standard editing symbols. It can also be corrected grammar, rhetoric, and paragraphs that are unclear. Usually people will continue to make the same syntax error, so it takes time to indicate. Next is a comment on organizational structure: Different degrees (clause, sentence, paragraph, subtraction or even chapter) order chaotic, redundant, independent content, and loss of losses. It is difficult to describe the characteristics of the comment on the content. You may suggest that the author expands its own ideas, considering a problem, error, potential problem, express praise, etc. "Because Y, you should read X" is a useful comment. You don't have to accept all your opinions, but you must take it seriously. It is quite sad, but often improves the level of the paper. You often find an opinion that there is a problem, but you feel unacceptable, then look for a third road. Publish more papers, which is actually easy than thinking. Basically, the standards for reviewing the thesis of AI publication reviews are: (a) a new meaning; (b) in certain aspects, in line with standards. Look at the IJCai's conference, you will find that the standard of the paper is quite low. This situation becomes worse due to the inherent randomness of the review process itself. So a paper published is to try. Ensure that the readability of the paper is better. The cause of the paper is rejected, in addition to meaningless, it is impossible to understand or organize bad. Prior to the pursuit of the journal, the paper should be exchanged for a period of time and make appropriate amendments based on feedback. It is necessary to resist the practice of rushing the results to the journal. In the field of Ai, there is no competition, and no matter what, the latency of the publishing cycle has much more time to comment on draft.
Read the journal or meetings of the conference you want to make sure that the style and content of your paper are suitable. Many publications have a "authors of the authors" and take a closer look. The main conferences will be awarded the award-winning papers that have been accomplished in the received papers and carefully study. It is usually an early report to which a relatively short part of the work content is submitted to the meeting, and then submit a final formal papers in the Journal. The thesis is determined - don't blaely. There is a big difference in the review process of journals and conferences. In order to save time, the review of the conference papers must be rapid, there is no time, and there is no time or exchange. If you are rejected, you have failed. But the journal papers are different, you can often argue with editorials, and argued by editing and reviewers. Reviewers generally help you. If you have received a more annoying review report, you should have a chairman or editorial complaint with the program of the General Assembly. You can't expect how much feedback can be obtained from the report of the conference paper review. But for journal papers, it is often a great suggestion. You don't have to do full according to the recommendations of the review report, but if you don't do it in accordance with the report, you must explain the reasons, and it is aware that this may result in further negative evaluation. Anyway, no matter which review, it is polite as a reviewer. Because in the rest of your career, you will be with the reviewers in an academic circle. Mit Ai Lab Memos is generally or close to the level. In fact, Technical Reports is basically a revision of these MEMOS. Working Papers is not official, this is a good way to divide your paper to colleagues. To publish these internal files, just receive a form to publications office, and have two faculty signed? Just like other research activities, the time spent by the paper is always better than expectations. High high. The publishing of the paper is more serious on this issue. When you have completed a papers, cast it out, wait for the publication. After a few months, the paper and comments were returned. You have to modify the paper. Then it is a few months before returning to your modification confirmation. If you have published the different forms of the paper, if there is a short investment, a long period of time, such a process will repeatedly round. As a result, it is possible that when you are tired, the research theme has also been tired of life, you are still modifying the paper. This reveals us: Don't do those who need enthusiastic investment but it is difficult to publish the papers - bitterness. 7. Another way to speak with peers exchange is the speech, the problem of writing the paper mentioned above, the same applies to speaking. Standing in front of the audience, it is very important to make the audience to sleep. It is very important to respect the recognition of others, respect and even the final job hunt. The ability of the speech is not born. Below is some ways to learn and practice: Patrick Winston has a very good text about how to do a speech. In January of each year, he will speak, demonstrate and describe its speech skills. If you feel that you are a bad speaker, or you want to be an excellent speaker, choose a public speech class. The primary performance class is also useful. If your mentor has a regular research semina, voluntarily gave a speech. Mit ai laboratory has a range of semicona symposium called Revolving seminar. If you think that some of your points is worthy of writing into AI Memo or in the context, you will be able to make a report. In-depth understanding of the different robot projects in the laboratory, when you come to friends and family, you can take them around and do a 60-minute report on the robot. Since the modified speech is far more likely to modify the paper, some people will feel that this is a good way to find ways to express my thinking.
(Nike Brady once again said that all of his best papers came from a speech). Practice in an empty room, the best is the report you want to do immediately. This helps to adjust the skills of the report: what is talked about each slide; the delay of the conversion and keeps smooth; maintain the synchronization of the interpretation and slide; the length of the estimated report. You spend the less time on the adjustment device, the longer the time with people who live with people, the more the mirror, recorders or recorder exercises is another method. The laboratory has these three devices. This also helps adjust your pronunciation and limb language. For comparative formal reports - especially your reply - should practice it in front of several friends, ask them to criticize. Observe how others do report. There are a lot of people who visit MIT will be reported. The report will be able to feel the area you are not familiar with, and if the report is not interested, you can secretly analyze where the reporter is wrong. Find a friend, tell him the nearest idea. This can improve communication skills, but also debugging your own ideas. 8. Programming is not all AI paper contains code, and many heavyweight characters in the art have never written an important program. But in order to initially approximate AI work principles, you must design it. Not only many AI research work needs to be written, and the learning programming can give you what can be calculated, which is the principal source of AI to cognitive scientific contribution. In MIT, all AI programming in nature uses Common Lisp. If you still don't know, hurry to learn. Of course, learning a language cannot be equated with the learning program design; some of the technologies included in the AI program design are different from those used in system programming or application design. When you start learning, you can first take a look at Abelson and Sussman's "Structure and Interpretation of Computer Program" and do some exercises. This book is not coheed with the AI program design, but contains some of the same technologies. Then read Winston and Horn written LISP book third edition, there are many elegant AI programs in the book. Finally, the actual programming is carried out, not reading, is the best way to learn. Learning LISP programming has many traditions. Some people are used to writing code, depending on the personality. There are also people looking for opportunities to learn from experienced programmers, or ask him to evaluate your code. Reading others' code is also a very effective way. If you can go to the senior classmate to ask for their source code. They may have some complaints about the bottom of the day, the drought, Tan Yizhen, 罴 蚴 蚴 蚴 蚴 喜 ⒉ ⒉ 芄ぷ髟 芄ぷ髟 2 2 2 还 还 还 创 创 创 芄ぷ髟 创 创 创 创 创 创 创 创 创 创 创 创 创 创 创 创 创 创 创 创 创 创 创 创 创 创 创 创 创 创 创 创 创 Xing Fu 粒 夂 夂 洹 洹 鹑 艘 艘 艘 艘 艘 艘 艘 艘 艘 艘 艘 艘 艘 艘 艘 艘 艘 艘 艘 艘 虼 艘 虼 艘 艘 艘 艘 虼 谀 谀 芬 龌蛘 龌蛘 龌蛘 鲅 鲅 鲅 说 鹑 淙ピ 淙ピ 淙ピ 氲 Crowning to rush to raise umbrella h 绻 蠖 蠖 蠖 蠖 豢 豢 斫 颍 崦 崦 ⑹ ⑹ ⑹ 颍 颍 颍 崦 崦 崦 崦 崦 崦 崦 崦 创 崦 创 创 创 创• Those knowledge learned in the software engineering class is still useful in the AI programming. To add a comment to the code. Use the correct data abstraction. Align the map and your code, because the language you use is basically Common Lisp, so the portability is very good. Such this. After a few years of study, you should write some own standard AI modules, such as true value maintenance system planner rules system different styles of interpreter have process analysis to optimize the compiler with inheritance characteristics Several search methods Based on Explain the learner anything you are interested in, you can try the implementation of the program. You can seize the essence of the problem, complete a function version in a few days.
Modifying existing procedures is another effective way, and the forefront of the front is awkward. It does understand the principle, advantages and disadvantages, and efficiency of its working principle. Unlike other usual programmers, there is little mutual borrowing code between AI programmers. (Demo code exception). This part is very effective due to the AI program. (Many famous AI programs work only on the three examples mentioned by the author's thesis, although this situation has improved recently). Another reason is that the AI program is often rushing to make together, and there is no considering of generalization problems. Using FOOBAR's "Standard" rule interpreter, it is very effective at the beginning, and will soon find that there are missing some features you need, or not enough efficiency. Although you can modify your own needs, remember to understand that others' code is very time consuming, sometimes it is better to write one. Sometimes the work that builds a standard package can become a paper. Like the paper, the program may be too pursuit of perfection. Do not locate the code for perfect, maximized abstract abstract, write macro and library, deal with the operating system core, which makes many people from their own papers and deviate from their own field. (From another aspect, perhaps this is what you need in the future). 9. Tutor in MIT, there are two types of mentor, teaching tutors and papers tutors. The work of teaching mentors is relatively simple. Every graduate student is allocated a teacher as a teaching faculty. The role of teaching instructor is the representative of the system, telling you what is the formal request for you, if your progress is slow, urge you, ratify your course plan. If everything goes well, you only need to see the teaching tutor twice a year, on the day of the registration day. On the other hand, if you have encountered difficulties, the teaching tutor will provide guidance to the system for you. The paper tutor is a person who supervises you. Choosing the paper tutor is the most important choice during your study, more important than the topic. In a broader sense, AI is learning through the master of the apprentice. There are many fields of technical or informal knowledge in the research process, can only be learned from the instructor, can't find in any textbook. Many AI majors are a weird person, and the graduates are the same. The relationship between the mentor and the graduate student is very personal, your personal characteristics must work well with the mentor, so that you can work together. Different mentors have different styles. Here are some factors that need to be considered: How much guidance do you need? Some tutors will give you a question that defines a good idea to make a papers, explains the solution and tell you how to work. If you are in a place, they will tell you how to carry out. Other tutors belong to the hand-type, they may have no help to your topic, but once you choose the topic, they have a very role in guiding your ideas. You need to consider clear your own independence or need guidance. How much do you need to contact? Some mentors have asked to meet you every week and listen to the report of your progress. They will tell you the paper you should read and give you practical exercises and projects. Other tutors will not exceed twice a semester. How much pressure do you can withstand? Some of the pressure applied to some mentor is very large. How much emotional support does it take? What is the seriousness of listening to mentor? Most tutors will be quite officially recommended for your papers. Some tutors are trustworthy, they give the recommendations, if implemented, almost certainly make an acceptable papers, if it is not an exciting papers. Others throw a lot of ideas, most of them are unrealistic, but there are some, perhaps leading to major breakthroughs. If you choose such a tutor, you first have to treat yourself as a filter.
What type of research group is provided? Some professors will create the environment and gather all students together, even if they are not the same project. Many professors will face their students every week or every two weeks. Is this useful for you? Can you get along with the professor's students? Some students have found them to build good work relationships with students from other teaching groups. Do you want to participate in a big project? Some professors will decompose a large system, and each student is part of it. This gives you an opportunity to discuss problems with a group of people. Do you want to supervise together? Some paper projects contain multiple AI fields, need you to establish a close work relationship with more than two professors. Although your official papers mentor only one, sometimes this does not reflect the actual situation. Do you want to guide the topic of the paper outside the field of research? Can you work with your instructor, more important than what you do itself. MIT's machines have guided the paper of quantum physics and cognitive modeling; instructors guiding visual papers. But some teachers will only guide their own fields in the field of interest. This is especially true for young teachers who want to get a lifelong position. Will the mentor fight for you? Some tutors will fight for you to fight in the system or some hostile entities. Sometimes the system is unfavorable to certain types of students (especially for female students and quirks), this is important. The mentor is willing to recommend your work at the meeting? This is part of the mentor work and is significant for your future. These factors described above, different schools are very different. MIT provides much free freedom compared to most schools. Looking for a paper tutor is the most important task of your graduate student's first grade. At the end of the study, or the secondary school year begins, you must have a papers tutor. Here are some of the tricks: check the research summary of the laboratory. One of these contends describes every teacher and what many graduate students are currently doing. If you are interested in the research work of some teachers, check the nearest papers. In the first semester, talk to as many teachers as possible. Go to feel what they like to do, what is their research and guiding style? Conversation with graduate students expected. To ensure multiple students with the mentor, because each mentor has different workfalls when communicating with different students, there are different works, and the 鼙 鲅 鲅 笥 笥 笥 笥 笥 笥 笥 笥 鲅 笥 鲅 笥 笥 鲅 鲅 鲅 鲅 鲅 涣 涣 涣 涣 鲅 鲅 鲅 鲅 餍 餍 餍 鲅 餍 餍 餍 鲅 鲅 鲅 鲅 餍 鲅 鲅 鲅 鲅 餍 鲅 鲅 鲅 鲅 鲅 鲅 鲅 鲅 鲅 鲅 鲅 鲅 鲅 鲅 鲅 鲅 鲅 鲅 鲅 鲅 鲅 鲅 鲅 笥 鲅 涣Many teachers' conferences are open to new students. This is a very good way to understand the way your mentor work. As a discipline, AI is a very common point is that many useful work is done by graduate students, not doctors - they are busy doing management. This has several consequences. One is a teacher's reputation, whether it will acquire life, to a large extent, depending on the student's work. This means that professors have a strong motivation to attract the best students to work, and give effectively guidance and sufficient support. Another consequence is that due to the direction of most student's paper is formed by the mentor, the direction and development of the entire field depends to a large extent on what kind of graduate students choose to choose. When the tutor is selected, determine the requirements of your tutor, make sure that the mentor knows. Don't waste time on the project you don't want to do due to bad communication. Don't completely depend on your mentor, to build your own network. It is very important to find some people who can regularly review your work, because it is easy to walk into the magic. People in the network can include research and teachers of their own laboratory or foreign units. In relationships with other students, teachers and even their own mentors, it is likely to encounter racists, gender discrimination, gay or other embarrassing things. If you missed, go seeking help. Mit's ODSA published a booklet called "Stop Harrassment", there are many suggestions. "Computer Science Women's Report" can be found in the LCS document room, also related to it. Some students in the laboratory are only guided by tutors in nominals.
This is very good for those who have strong independence. But if you have completed the work of an instructor, unless you have no tutor, you have a firm support network, don't do it. 10. The paper will take the graduation papers to occupy most of the postgraduate life, mainly to study, including topics, which is more than the actual writing. The purpose of the master's thesis is to make a doctoral paper training. Doctoral level research If it is not ready, it is difficult to do. The most nature of master's thesis is to show your master: You have fully understood the latest developments in this field and have the appropriate level of operation. Don't need to expand the latest knowledge in this field, and don't ask for your papers. However, our laboratory papers is always atmospheric, so many master's thesis actually make a significant contribution to the development of the art, about half of them published. This is not necessarily a good thing. Many people have concentrated on master's work, so Mit has a reputation: the quality of master's thesis is often higher than soldiers. This is contrary to the master's degree in the original purpose of preparing for Ph.D. Another factor is that the research has to contribute to the field, at least two years, which makes the graduate study time is unbearable. Now you may not feel a hurry, but when you have stayed in the laboratory, you must not escape from it. Master's degree in graduation is two and a half years, but the computer system strongly encourages students graduated in advance. If a master's topic is too large, you can decompose, partially to do a master's thesis, and the other is divided into doctoral thesis. Want to understand what is the study of master's thesis, read a few newest master's thesis. Remember that a better paper is the publishing or becoming a technical report, because this marks that the paper is considered to be extended to the latest knowledge - in other words, their papers are far beyond the level of master's thesis. Also read some papers that pass, without publishing, all pass the papers can be found in the MIT Library. Doctoral thesis must expand the latest knowledge, and the research of doctoral thesis must have the quality of published. MIT's temperament has been manifested, and many doctoral thesis have been authoritative in a certain field in a few years. For MIT's doctoral paper, create a new field, or put forward and solve a new problem, is not a great thing. Although this is not required. In general, it takes two to three years to do doctoral thesis. Many people spend more than two years to say goodbye and topic. During this time, you can try some things, such as taking assist towards or to lay a solid foundation or organize a team in a non-AI field. The actual writing time of doctoral thesis is about one year. The topic is the most important part of the paper work: Good papers can not only express personal views, but also communicate with peers. Choosing the topic must be you willing to pour enthusiasm. Personal vision is the reason for you as a scientist is your most concerned, principles, ideas or goals. There are a variety of forms. Perhaps you want to create a computer that can talk to it, maybe you want to save human beings from your computer's stupid use, maybe you want to exhibit all things to be unified, maybe you want to find new lives in space. Vision is always bigger, your paper does not realize your vision, but can work towards the direction. When doing papers, the most difficult thing is how to reduce the problem to a resolved level, while scale is a paper. "Solving the width of AI" is an example of common problems, and the topic is too great. You will find the range of constant reduction questions. The topic is a gradual process, not a discrete event, will continue until you have completed the paper. In fact, solving problems is usually much easier than accurately describing problems.
If your goal is a 50-year engineering, what is the reasonable ten-year project, one year? If the structure of the target is huge, what is the core component, how to understand the core part? An important factor is how much the risk can be tolerated. Weighing between the final success and the risk. This is not always right, there are many researchers who have not involved in AI. A good passion has a central part. You are sure that you can do it, and you and your mentor agree that this has met the graduation requirements. In addition, there are also many expansions in the paper, there is a possibility of failure, but if it is successful, it will increase the wonderfulness of the paper. Although not every paper selection is compliant, it is worth a try. Some people think that work in multiple projects can be done when choosing questions. This really reduces the risk. Other people are willing to choose a separate question before doing any work. Maybe you are only interested in a certain area so that your topic is much narrow. Sometimes, you will find the teacher's teacher without a person to guide you. Maybe it will also find that there is no natural topic in that field, but there is a good idea for other fields. Master's topic is more difficult than doctoral topics, because master's thesis must be completed when you know not much confidence. A factor in the doctor needs to consider whether to continue the fields studied in the master's phase, may expand or be found, or simply go to another area. Things to be in the same area is simple, it may take only one to two years, especially if the topic for doing doctoral text has been discovered in the work of the master's phase. The deficiency is that it is easy to set up, and the field can increase the width of knowledge. Some papers are very novel, and some are very ordinary. The former has created a new field and explored the phenomenon that had not been studied before, or provides an effective solution for difficult problems; the latter is perfectly solved the definition of good problems. Both papers are valuable. Which papers are selected depend on the personal style. The "future work" part of the paper is a good source of papers. No matter what kind of topic, it must be that the former has not been done. Even if someone works at the same time, it is not good. There are a lot of things to do, there is no need for competition. There is also a common situation. After reading the papers of others, I feel very panic, as if it has already solved your problem. This usually occurs during determining the problem of the paper. In fact, it is only similar to the surface, so gives the paper to a high person who knows your work, see what he said. Mit Ai Laboratory's paper is not all relevant to artificial intelligence; some is related to hardware or programming languages. After choosing a good question, even if you have a little virtual, you have to answer the following questions: What is the argument of the paper? What do you want to clarify? You have to have a saying, a five-minute answer. If you don't know what you are doing, others will not treat your topic, worse, you will fall in the selection - the circle of the topic, you can't extricate themselves. After starting to study the research, you must be able to interpret each part of the theory and implementation of each part. Remember, once the topic is selected, you must have a clear consistency with the standard of the instructor to complete the paper. If you have different expectations with him, you will definitely die very badly. It must be defined as a standard that completes tests, like a series of examples that prove your theory and procedures. This is a must do, that is, your mentor does not require this. If the environment has undergone fundamental changes, the test also changes. First try the simplified version of the issue. Use an example test. Before formation theoretical abstraction, a complete exploration has a representative example. There are many ways to waste time during the process of doing the paper.
To avoid the following activities (unless it is true to be related to the paper): the design of language expression; the user interface or the graphical interface is excessive; inventive new formal method; excessively optimized code; create a tool; bureaucratic style. Any work that is not very relevant to your paper should be reduced as possible. A well-known phenomenon "Thesis Escape" is that you suddenly discovered that the bug that corrected a operating system is very attractive and important. At this point you always consciously unconsciously deviate from the work of the paper. To remember what you should do. (This article is a paper escape from the paper for some authors). 11. Research Methodology This part is relatively small, please add research methods to define what is research activities, how to conduct research, how to measure research, and what is successful. AI's research methodology is a hodgepodi. Different methods are defined different research schools. The method is the tool. Use it, don't let them use you. Don't take yourself in the slogan: "Ai research needs to be firm", "philosophers will only talk about it, and artificial intelligence needs to work." "Before asking why, what is the calculation is calculated." In fact, it is necessary to succeed in the artificial intelligence, you have to be good at various technical methods, and you must also have a skeptical attitude. For example, you must be able to prove the aimation, and you must also think about whether the aizer explains what. Many excellent AI chapters are uniform in several ways to be balanced. For example, you have to choose a best route between too many theories (possibly unrelated to any practical issues) and cumbersome implementation (to express the actual solution to the expression "). You often face research decisions from "clean" and "dirty". Should you take time to give your questions to some extent? Still maintain the original state of the problem, at this time, although the structure is poor, it is close to the actual situation? The previous method (if a feasible) will be clearly determined, but this process is often cumbersome, or at least not directly solving the problem. The latter has the risk of caught in the vortex of various processes. Any work, no one must make a wise balance. Some work icons science. How do you observe how people learn arithmetic? How is the brain work, how to jump in the kangaroo, then engage in principle, form a testing theory. Some work icons project: strive to create a better question ruler or algorithm. Some jobs like mathematics: deal with formal deal, to understand attributes, give prove. Some work is the instance driver, and the goal is to explain the specific phenomenon. The best work is the combination of above. Method has social, see how others overcome similar problems, ask others how to handle some special case. 12. Emotional factors research are hard work and it is easy to lose interest. A embarrassing fact is that students who read Loofeng in this laboratory have only a few proportion of the final degree. Some people leave because they can earn more money in the industry, or because of individual reasons; the most important reason is due to the paper. The goal of this section is to explain the cause of this situation and give some beneficial recommendations. All studies have risks. If your project is impossible to fail, it is developing, not research. How hard is it for the failure of the project, it is easy to explain your responsible project failure as your own failure. Although this is actually proves that you have courage to challenge difficulties. There are very few people in the artificial intelligence area always have been successful, and the paper is out of the past. In fact, failure is often. You will find that they often do a few projects at the same time, only some is successful. The final successful project may have repeatedly failed. After many after the failure of the method, it took the final success. In your future career, it will experience a lot of failure.
But every failure project represents your work, many ideas, thinking, and even prepared code, you have found that you can use another completely different project after several years. This effect only appears after you have accumulated a considerable failed failure. Therefore, there will be a belief that will work after the initial failure. The actual time spent on the study is often much more than the plan. A small trick is to assign three times the expected time to each child task (some people add a sentence: "... even considering this principle"). The key to success is to make research becoming part of your daily life. Many breakthroughs and inspiration happen when you walk. If you think about it, you will find the Spring Festival. Successful AI researchers, persistent, is generally greater than the heroes. "Try" is also very important, which is the ability to distinguish the shallow and important ideas. You will find that your proportion of success is random. Sometimes, I have done the job that I need three months before I have done it. This is a delightful, making you more willing to work in this field. Others, you are completely trapped, I can't do anything. This situation is difficult to handle. You will feel that you will never do anything worth anything, or feel that you will no longer have the quality of the researchers. These feelings are almost definitely wrong. If you are a student admitted by Mit, you are absolutely qualified. You need to pause, maintain a high degree of tolerance for bad results. You have a lot of work to do through a short-term goal in regular settings, such as weekly or monthly. There are two ways to increase the possibility of achieving these goals, you can record the target in the notebook and tell another person. You can agree with a friend to exchange weekly goals and see who eventually achieves their goals. Or tell your mentor. Sometimes you will completely trap, similar to the idea of writing process, there are many possible reasons, but there is no solution. The range is too broad, you can try to solve the child problems in the process. Sometimes doubts about your research capabilities will throw away all your enthusiasm and make you no matter. It is necessary to keep in mind that the ability to learn is learning, not born. If you find yourself into a serious dilemma, there is no progress in a week, trying to work only one hour a day. A few days later, you may find everything back to the right track. Fear failure will make research work more difficult. If you find yourself unable to complete, ask if you have your own ideas due to an evade experiment. It is found that this time in the last few months is completely in the bill, it will stop you from further work. There is no way to avoid this, as long as you recognize failures and waste is part of the research process. Look at the book of Alan Lakien "How to Get Control of Your Time and Your Life", which contains many invaluable methods that enable you to enter the creative status. Many people find their own personal life and research ability. For some people, when everything is not as good, work is a refuge. Other people can't work if life is in chaos. If you feel that you really have a sad self-extricted, go see a psychologist. A informal survey showed that about half of our laboratory had seen a psychologist during the study period. One reason why artificial intelligence is so difficult is not universally accepted success standards. In mathematics, if you have a certain aimation, you do have something to do; if you don't have anyone else, then your job is exciting. Artificial intelligence has borrowed some standards from related disciplines, as well as some standards. Different practitioners, subsequent areas and schools will emphasize different standards. MIT is more emphasized than other schools, but there is also a big difference within the laboratory. One consequence is that you can't make all people satisfaction. Another consequence is that you can't determine if you have made progress, which will make you feel unsafe.
