Next: Endnote Previous: Research methodology Up: How To Do Research In the MIT AI Lab
Research is hard. It is easy to burn out on it. An embarrassingly small fraction of students who start PhD programs in AI finish. AT MIT, almost all those who do not finish drop out voluntarily. Some leave because they can make more money in industry, or for personal reasons; the majority leave out of frustration with their theses. This section tries to explain how that can happen and to give some heuristics that may help. Forewarned is forearmed: mostly it's useful to know that the particular sorts of tragedies, aggravations, depressions and triumphs you go through in research are necessary parts of the process, and are shared with everyone else who does it.
All research involves risk. If your project can't fail, it's development, not research. What's hard is dealing with project failures. It's easy to interpret your project failing as your failing; in fact, it proves that you had the courage to do something difficult.
The few people in the field who seem to consistently succeed, turning out papers year after year, in fact fail as often as anyone else. You'll find that they often have several projects going at once, only a few of which pan out. The projects that do succeed have usually failed repeatedly, and many wrong approaches went into the final success.
As you work through your career, you'll accumulate a lot of failures. But each represents a lot of work you did on various subtasks of the overall project. You'll find that a lot of the ideas you had, ways of thinking, even often bits of code you wrote, turn out to be just what's needed to solve a completely different problem several years later. This effect only becomes obvious after you've piled up quite a stack of failures, so take it on faith as you collect your first few that they will be useful later.
Research always takes much, much longer than it seems it ought to. The rule of thumb is that any given subtask will take three times as long as you expect. (Some add, `` even after taking this rule into account.'')
Crucial to success is making your research part of your everyday life. Most breakthroughs occur while you are in the shower or riding the subway or windowshopping in Harvard Square. If you are thinking about your research in background mode all the time, ideas will just pop out. Successful AI people generally are less brilliant than they are persistent. Also very important is ``taste,'' the ability to differentiate between superficially appealing ideas and genuinely important ones.
You'll find that your rate of progress seems to vary wildly. Sometimes you go on a roll and get as much done in a week as you had in the previous three months. That's exhilarating; it's what keeps people in the field. At other times you get stuck and feel like you can't do anything for a long time. This can be hard to cope with. You may feel like you'll never do anything worthwhile again; or, near the beginning, that you don't have what it takes to be a researcher. These feelings are almost certainly wrong; if you were admitted as a student at MIT, you've got what it takes. You need to hang in there, maintaining high tolerance for low results.
You can get a lot more work done by regularly setting short and medium term goals, weekly and monthly for instance. Two ways you can increase the likelihood of meeting them are to record them in your notebook and to tell someone else. You can make a pact with a friend to trade weekly goals and make a game of trying to meet them. Or tell your advisor.
You'll get completely stuck sometimes. Like writer's block, there's a lot of causes of this and no one solution.
Setting your sights too high leads to paralysis. Work on a subproblem to get back into the flow.
You can get into a positive feedback loop in which doubts about your ability to do the work eat away at your enthusiasm so that in fact you can't get anything done. Realize that research ability is a learned skill, not innate genius.
If you find yourself seriously stuck, with nothing at all happening for a week or more, promise to work one hour a day. After a few days of that, you'll probably find yourself back in the flow.
It's hard to get started working in the morning, easy to keep going once you've started. Leave something easy or fun unfinished in the evening that you can start with in the morning. Start the morning with real work-if you start by reading your mail, you may never get to something more productive.
Fear of failure can make work hard. If you find yourself inexplicably ``unable'' to get work done, ask whether you are avoiding putting your ideas to the test. The prospect of discovering that your last several months of work have been for naught may be what's stopping you. There's no way to avoid this; just realize that failure and wasted work are part of the process.
Read Alan Lakien's book How to Get Control of Your Time and Your Life, which is recommended even by people who hate self-help books. It has invaluable techniques for getting yourself into productive action.
Most people find that their personal life and their ability to do research interact. For some, work is a refuge when everything else is going to hell. Others find themselves paralyzed at work when life is in turmoil for other reasons. If you find yourself really badly stuck, it can be helpful to see a psychotherapist. An informal survey suggests that roughly half of the students in our lab see one at some point during their graduate careers.
One factor that makes AI harder than most other types of work is that there are no generally accepted standards of progress or of how to evaluate work. In mathematics, if you prove a theorem, you've done something; and if it was one that others have failed to prove, you've done something exciting. AI has borrowed standards from related disciplines and has some of its own; and different practitioners, subfields, and schools put different emphases on different criteria. MIT puts more emphasis on the quality of implementations than most schools do, but there is much variation even within this lab. One consequence of this is that you can't please all the people all the time. Another is that you may often be unsure yourself whether you've made progress, which can make you insecure. It's common to find your estimation of your own work oscillating from ``greatest story ever told'' to ``vacuous, redundant, and incoherent.'' This is normal. Keep correcting it with feedback from other people.
Several things can help with insecurity about progress. Recognition can help: acceptance of a thesis, papers you publish, and the like. More important, probably, is talking to as many people as you can about your ideas and getting their feedback. For one thing, they'll probably contribute useful ideas, and for another, some of them are bound to like it, which will make you feel good. Since standards of progress are so tricky, it's easy to go down blind alleys if you aren't in constant communication with other researchers. This is especially true when things aren't going well, which is generally the time when you least feel like talking about your work. It's important to get feedback and support at those times.
It's easy not to see the progress you have made. ``If I can do it, it's trivial. My ideas are all obvious.'' They may be obvious to you in retrospect, but probably they are not obvious to anyone else. Explaining your work to lots of strangers will help you keep in mind just how hard it is to understand what now seems trivial to you. Write it up.
A recent survey of a group of Noble Laureates in science asked about the issue of self-doubt: had it been clear all along to these scientists that their work was earth-shattering? The unanimous response (out of something like 50 people) was that these people were constantly doubting the value, or correctness, of their work, and they went through periods of feeling that what they were doing was irrelevant, obvious, or wrong. A common and important part of any scientific progress is constant critical evaluation, and is some amount of uncertainty over the value of the work is an inevitable part of the process.
Some researchers find that they work best not on their own but collaborating with others. Although AI is often a pretty individualistic affair, a good fraction of people work together, building systems and coauthoring papers. In at least one case, the Lab has accepted a coauthored thesis. The pitfalls here are credit assignment and competition with your collaborator. Collaborating with someone from outside the lab, on a summer job for example, lessens these problems.
Many people come to the MIT AI Lab having been the brightest person in their university, only to find people here who seem an order of magnitude smarter. This can be a serious blow to self-esteem in your first year or so. But there's an advantage to being surrounded by smart people: you can have someone friendly shoot down all your non-so-brilliant ideas before you could make a fool of yourself publicly. To get a more realistic view of yourself, it is important to get out into the real world where not everyone is brilliant. An outside consulting job is perfect for maintaining balance. First, someone is paying you for your expertise, which tells you that you have some. Second, you discover they really need your help badly, which brings satisfaction of a job well done.
Contrariwise, every student who comes into the Lab has been selected over about 400 other applicants. That makes a lot of us pretty cocky. It's easy to think that I'm the one who is going to solve this AI problem for once and for all. There's nothing wrong with this; it takes vision to make any progress in a field this tangled. The potential pitfall is discovering that the problems are all harder than you expected, that research takes longer than you expected, and that you can't do it all by yourself. This leads some of us into a severe crisis of confidence. You have to face the fact that all you can do is contribute your bit to a corner of a subfield, that your thesis is not going to solve the big problems. That may require radical self-reevaluation; often painful, and sometimes requiring a year or so to complete. Doing that is very worthwhile, though; taking yourself less seriously allows you to approach research in a spirit of play.
There's at least two emotional reasons people tolerate the pain of research. One is a drive, a passion for the problems. You do the work because you could not live any other way. Much of the best research is done that way. It has severe burn-out potential, though. The other reason is that good research is fun. It's a pain a lot of the time, but if a problem is right for you, you can approach it as play, enjoying the process. These two ways of being are not incompatible, but a balance must be reached in how seriously to take the work.
In getting a feeling for what research is like, and as inspiration and consolation in times of doubt, it's useful to read some of the livelier scientific autobiographies. Good ones are Gregory Bateson's Advice to a Young Scientist, Freeman Dyson's Disturbing the Universe, Richard Feynmann's Surely You Are Joking, Mr. Feynmann!, George Hardy's A Mathematician's Apology, and Jim Watson's The Double Helix.
A month or two after you've completed a project such as a thesis, you will probably find that it looks utterly worthless. This backlash effect is the result of being bored and burned-out on the problem, and of being able to see in retrospect that it could have been done better-which is always the case. Don't take this feeling seriously. You'll find that when you look back at it a year or two later, after it is less familiar, you'll think ``Hey! That's pretty clever! Nice piece of work!''
A whole lot of people at MIT