"Reprint" MIT AI Lab: How do you do research?

From the MIT AI Lab: How do I do research?

All graduate students in artificial intelligence laboratory

Editor:David Chapman

Version:1.3

Duration:1988 Year 9 months

Translator: Liu School of Information, Beijing Normal University, PhD

Copyright 1987, 1988 author All rights reserved

The main thrust of this article is to explain how to do research. The advice we provide is invaluable for doing research itself (reading, writing and programming), understanding the research process, and beginning to love Research (methodology, selection, mentoring, and affective factors) .

Note: The AI Lab's working Papers is used for internal communication and contains information that is too preliminary or too detailed and cannot be published. Unlike a formal paper, all references are listed.

1. Introduction

What is it?

There is no God Dan Elixir can guarantee success in the study, this article just lists some of the informal comments that may be helpful.

Who is the target audience?

This document is written primarily for new graduate students enrolled in the MIT AI Lab, but it is also valuable for AI researchers in other organizations. Even if you are not a researcher in the field of artificial intelligence, you can find a valuable part of yourself.

How do I use it?

To read this article, too long, it is best to use the way of browsing. Many people feel that the following method works well: read through it quickly, and then select the part that is relevant to your current research project.

This document is roughly divided into two parts. The first part deals with the various skills that researchers need to have: reading, writing and programming, and so on. The second part discusses the research process itself: how to study what is going on, how to do research, how to select the subject and choose the Tutor, how to consider the emotional factors in the study. Many readers have reflected that the second part is more valuable and more interesting in the long run than the first part.

Section 2 How to read a good foundation for AI Research. Lists important AI journals and gives some tips on how to read them.

Section 3 How to become a part of AI Research: Keep in touch with people who can keep you on the forefront of research and know what materials to read.

Section 4 Learn about AI -related areas of knowledge. Have a basic understanding of several areas and be proficient in one or two areas.

Section 5 How to do research notes.

Section 6 How to write journal papers and graduation papers. How to write reviews for drafts, and how to make use of other people's assessment opinions. How to publish a paper.

Section 7 How to do the research report.

Section 8 is about program design. AI programming differs from the usual programming.

Section 9 The most important questions about the study career, how to choose the tutor. Different instructors have different styles, and the advice in this section will help you find the right mentor. A mentor is a resource that you must know how to use.

Subsection of the thesis. The graduation thesis will occupy most of the postgraduate career, this part involves how to select the topic,

And how to avoid wasting time.

The study methodology has not been completed.

Perhaps the most important section: the emotional factors involved in the research process, including how to face failure, how to set

Goal, how to avoid insecurity, maintain self-confidence, enjoy happiness.

2. Read

Many researchers spend half their time reading the literature. Many things can be learned quickly from the work of others. This section discusses reading in AI , and the fourth section discusses other topics related to reading.

Read the literature, beginning today. Once you start writing a paper, there is not much time, and then the reading is focused on the literature of the topic of the thesis. In the first two years of graduate school, most of the time is spent doing coursework and laying the groundwork. At this point, read the textbook and published journal articles. (Later, you will read the draft of the article mainly, see Section Three).

The amount of reading required to lay a solid foundation in this area is prohibitive. But since AI is only a small area of research, you can still spend a few years reading the most essential part of a large number of papers published in this field. A useful trick is to find out the most essential papers first. Here are some useful books to look at: for example, a list of recommended readings for postgraduate programs in other schools (mainly Stanford), which can give you some initial impressions. If you are interested in a sub-area of AI , ask a senior graduate in this field what the 10 most important papers in this field are, and if so, borrow it to copy it. Recently, there have been a number of carefully edited essays on a sub-domain, especially published by Morgan-kauffman .

AI Labs has three in-house publication series:working Papers,memos and Technical Reports, and the formal level is increased in turn And can be found on the shelves on the eight floor. Review the publications of recent years and copy them with great interest. This is not only because many of them are significant papers, but also important for understanding the progress of the work of laboratory members.

There are a lot of journals about AI , and fortunately, only a few are worth seeing. The core issue is Artificial Int .

elligence, there are also writing"The Journal of Artificial Intelligence"Or"AIJ"OfAIThe real value of the papers in the field will eventually go toAIJ, so it's worth browsing every year of each periodAIJBut the journal also has a lot of papers that make people upset.Computational Intelligenceis another journal worth reading.cognitive Science also publishes many significant ai thesis. machine learning is the most important resource in the field of machine learning. ieee pami (pattern analysis and Machine Intelligence ) is the best visual periodical with two or three valuable papers in each issue. IJCV ) was newly founded and has been valuable so far. robotics the article is mainly about dynamics, Sometimes there are also epoch-making intelligent robot papers. ieee Robotics and Automation occasional good articles.

Every year, you should go to your school's computer Science library (at MIT 's tech Square ), read the AI Technical report published by other colleges, and choose what you are interested in reading carefully.

Reading a paper is a skill that needs to be practiced. It is impossible to read all the papers in full. Reading papers can be divided into three stages: the first stage is to see if there is something interesting in the paper. AI papers contain summaries, which may have content, but they may not be well summarized, so you need to jump in and see a little bit and see what the author has done. Content directory (thetable of Contents), the concluding section (conclusion), and introduction (Introduction ) is a three focus. If none of these methods work, then you have to go through the sequence quickly.

Once the approximate and innovative points of the paper are made clear, it is possible to decide whether a second phase is needed. In the second stage, find out what part of the paper really has content. Many pages of paper can be rewritten to a page or so, so you need to look for places that are really exciting, often hidden somewhere. Where the author of the paper finds interest in his work, it may not be of interest to you, and vice versa. Finally, if you feel that the paper is indeed valuable, go back to the entire intensive reading.

One question to keep in mind when reading a paper is, "How should I use this paper?" "Is it really like the author claims?" "" If ... What's going to happen? ”。 It is not the same as understanding the thesis to understand what the paper has been drawn. Understanding the paper, you

To understand the purpose of the paper, the author's choice (many are implicit) hypothesis and formalization is feasible, the paper points out

What direction, what are the problems in the field of the thesis, what are the persistent difficulties in the author's research,

What are the strategic views expressed in the paper, and so forth.

It is helpful to associate reading with programming. If you are interested in a particular area, after reading some papers, try to implement the "toy" version of the program described in the paper. This will undoubtedly deepen understanding.

Sadly, many AI Labs are inherently withdrawn, with members reading and quoting their own school labs. Be aware that other institutions have different ways of thinking about problems, are worth reading, take it seriously, and cite their work, even if you think you are aware of their mistakes.

Often someone will hand you a book or a paper and tell you that you should read it because there is a very shiny place and / or it can be applied to your research work. But when you're done reading, you find nothing particularly shiny, just barely available. So, the confusion came, "I what is wrong ah?" Did I miss something? ”。 In fact, this is because your friend, in reading a book or paper, is catalyzed by some of the ideas that have been formed in his mind, to see where it is valuable to your research project.

3. Building Relationships

A year or two later, there were some ideas about the sub-areas that I was prepared to engage in. At this point--or before one o'clock-- It is important to join the Secret Paper passing Network . This informal organization is a reflection of what artificial intelligence is really doing. The work leading up to the trend will eventually become a formally published paper, but at least one year after the Bulls fully understand it, the bulls are at least one year ahead of the new Idea's work.

How do cattle people find new ideas? May be heard from a meeting, but most likely fromSecret Paper passing Network。 Here's a general picture of how the network works.Jo CoolHave a good idea. She merged the incomplete implementation with some other work and wrote a draft paper. She wanted to know what the idea was, so she sent a copy of the paper to 10 friends and asked them to comment. Friends thought it was a great idea and pointed out the mistakes, and then the friends copied the papers to their friends and continued. A few months later,jo has been extensively revised and sent to < Span style= "font-family: ' Times New Roman ';" >aaai . Six months later, the paper was formally published in five pages (this is aaai the length of the proceedings allowed). Finally jo began to tidy up related programs and wrote a longer paper (based on aaai published the feedback from the paper). Then sent to ai Journal. ai the journal takes about two years to review the paper, including the time it takes for the author to revise the paper, and the corresponding publication delay. Therefore, ideally, jo 's idea of final publication in the journal takes about three years. So it's too late for cows to learn anything from journal articles published in the field.

You, too, can be a bull man. Here are some tips for building an academic network: There are a lot of mailing lists that discuss an AI sub-domain, such as connectivity or vision, and choose the list of interests you want to join. When discussing your thoughts with people who are familiar with the field, they may not directly evaluate your thoughts, but rather say, "Have you ever read XXX?" "It's not a question, it's a suggestion that you read a document and it's probably related to your idea." If you haven't read the literature, get detailed information about the document from the expert you're talking to, or borrow a copy from him directly.

When you read a paper that excites you, five copies are sent to the other five who are interested in it. They might get back good advice.

The lab has a lot of informal (sustainable) paper discussion groups for different sub-areas, and they get together once a week or every two weeks to discuss the papers they have read.

Some people don't mind looking at their desks, that is to say, by flipping through the papers they're piling up on their desks that they'll be reading or often flipping through. You can go over it and see if you are interested. Of course, first of all to get the master's permission, to know that some people really dislike others to turn their own things. Try those people who are approachable.

Also, some people don't mind you looking through their file cabinets. There are many knowledgeable people in the lab, and there are a lot of treasures in their file cabinets. This is often a quicker and more reliable way of finding papers than using a school library.

Once you have written something, distribute the copy of the draft to those who may be interested. (There is also a potential problem: Although there is little plagiarism in the AI field, it does.) You can write "Please do not photocopy or quote" on the first page to make some precautions. Most people don't read most of the papers they receive, so don't worry if only a few of them return to comment. You can do this a couple of times--this is necessary for a journal paper. Note that, in addition to their own mentors, it is generally rare to send more than two drafts to the same person.

When you have finished writing a paper, copy the paper to those who may be interested. Don't assume that people will naturally read journals or conferences that publish papers. It is even more difficult to read in-house publications (memos and technical reports).

The more people you keep in touch with, the better the effect. Try to exchange papers with people from different research groups,AI Labs, and different academic fields. Make yourself a bridge between the two research groups that are not connected, so that soon your table will have a stack of related papers.

a reference b and b reference c and d , c references d , and so on. Note Those papers that are often quoted, which are usually worth reading. The reference map has wonderful properties. One is that research groups that often have research on the same subject do not understand each other. You search for the graph and suddenly find a way to get into another part, which usually occurs in different schools or in different ways. It is valuable to know as many methods as possible, which is better than a very deep understanding of one approach.

Temporarily shelved. Talk to others. Tell them what you are doing and ask what they are doing. (If you're shy about talking to other students about your ideas, keep talking, even if you don't have any ideas, talk to them about a paper you think is really good.) This will naturally lead to a discussion of what to do next. There is an informal lunch seminar on the seven floor of the activity building every noon. In our lab, people are accustomed to working at night, so it's possible to have a loose-knit group of people for lunch.

If you have a lot of communication with the outside world-doing presentations or attending meetings-going to sheet business cards, it's important to make your name easy to remember.

From a certain time onwards, you will begin to attend an academic conference. If you do participate, you will find the fact that almost all of the conference papers are boring or stupid. (The reasons for this are interesting, but not related to this article, not to be discussed). Why did you go to the meeting? The main purpose is to meet people outside the laboratory. People outside will disseminate news about your work, invite you to report, tell you about the academic ethos and characteristics of the researcher, introduce you to other people, help you find a summer job, and so on. How to get acquainted with others? If you think someone's paper is valuable, run up and say, "I appreciate your paper very much" and ask a question.

Gain access to other labs for summer jobs. This way you will meet another group of people and perhaps learn another way to look at things. Ask Seniors how to get this opportunity, they may already have worked where you want to go, and can help you connect.

"Reprint" MIT AI Lab: How do you do research?