CMU老教授对Ph.D的理解

多年前看到的CMU教授写给自己学生的一篇关于如何读Ph.D的文章，甚是经典，特转来与诸位共勉。

USEFUL THINGS TO KNOW ABOUT PH.D. THESIS RESEARCH
H. T. KUNG

(Prepared for "What is Research" Immigration Course,
Computer Science Department, Carnegie Mellon University, 14 October 1987)

Presntation Outline

1. Introduction
2. Why Ph.D. thesis could be really difficult for a student
3. Types of Ph.D. theses (from Allen Newell)--not a topic of this talk
4. Growth of a star (the transformation process that some students go
through to become a mature researcher)--which stage are you in?
5. Stages of Ph.D. thesis research
6. Methods to get into the depth of a topic (or how to come up with
good ideas)
7. Breaking myths
8. Pitfalls to avoid (easy ones to avoid listed first)
9. Some other general advice
10. All the effort is worth it (believe it or not)

1. Introduction

- Ph.D. thesis is treated very seriously at leading universities.

* Expectation is high.

- Ph.D. thesis represents a substantial work. Faculty
often tell other people that "We have a student
working on this area for his or her Ph.D. thesis."
Amazingly enough, this is usually sufficient to
convince people that the problem is somehow going to
be solved.

* Ph.D. thesis research is a task to ensure that the student
can later take on independent, long-term research
commitments. (If a Ph.D. student does not intend to be a
researcher, the Ph.D. thesis work is not worth the effort
in general at least at CMU.)

* Through the Ph.D. thesis process the student is
transformed into a professional researcher.

* Faculty are judged by the theses of their Ph.D. students.

* High standard Ph.D. thesis is probably one of the most
important factors that contribute to the success of
graduate education at leading American universities.

* Ph.D. thesis is probably the only real challenge for
getting a Ph.D. degree.

- Ph.D. qualifier is seldom a problem for motivated
students.

- Ph.D. thesis research is probably more mechanical than a new
graduate student would think. (Of course the process is still
too complex to be automated.)

* Knowing this mechanism can be more important than thesis
results themselves.

* Some information presented here may be relevant to your
whole research career, i.e., it is not just for the Ph.D.
thesis per se.

- This talk consists of pragmatic advice.

* The talk is based on my personal experience (i.e., not
based on any serious research)

- I happen to have research experience in both theory
and system areas. We will compare thesis research in
these two areas.

* This is a common sense talk and will have down to earth
discussions.

- "I wish someone told me this before."

2. Why Ph.D. thesis could be really difficult for a student

- Most likely this is your first, major research experience.

* A big challenge for most students

- No simple recipe

* Different talents

* Different kinds of theses

* Different approaches

- The work is judged by thesis committee (mostly advisor). This
produces anxiety.

* Unlike other research you will do, the evaluation
mechanism for thesis research is very unique.

* No clear contract

* No clear standard (we only know it is high)

* Recall the Stanford murder case (the former student said,
after he had finished--he did finish something-- his jail
term, that he might do it again under a similar
circumstance).

3. Types of Ph.D. theses (from Allen Newell)--not a topic of this talk

- Opens up new area

- Provides unifying framework

- Resolves long-standing question

- Thoroughly explores an area

- Contradicts existing knowledge

- Experimentally validates theory

- Produces an ambitious system

- Provides empirical data

- Derives superior algorithms

- Develops new methodology

- Develops a new tool

- Produces a negative result

4. Growth of a star (the transformation process that some students go
through to become a mature researcher)--which stage are you in?

- Knowing everything stage

* Student: "I have designed a supercomputer even before
graduate school."

* Faculty: speechless

- Totally beaten up stage

* Student: speechless

* Faculty: smiling at the student's progress so
communication is possible now.

- Confidence buildup stage

* Student: "I am not stupid after all." (student thinks)

* Faculty: "Uh oh, she is ready to argue." (faculty think)

- Calling the shot stage

* Faculty: "I am going to design an n-processor
supercomputer."

* Student: "You are crazy, because ..."

5. Stages of Ph.D. thesis research

a. Selection of area--not a topic of this talk

b. Selection of advisor--not a topic of this talk

c. Becoming a researcher in the area

- Building up general knowledge, experience, and confidence

- Knowing issues and important questions in the area

- Capturing research opportunities

* Don't let any idea or question go by without first
giving it careful thought.

- Be alert and diligent.

* Pay attention to new technologies

- Examples

* VLSI, networking, and new chips such as
the Weitek floating-point chips three
years ago which in some sense gave the
initial motivation for the Warp project

- Some useful things to do (from Dave Gifford, MIT)

* Read recent proceedings of the best conferences, and
ask more senior people what were the best papers.
Try to figure out what makes a great paper (and thus
what makes great research).

* Keep a notebook that contains your research notes.
Put all of your empirical data and initial ideas in
the notebook. Make notes on a paper as you read it
and think about the assumptions of the author and
the importance of the results.

* Follow references from one paper to another until
you know an area extremely well. Don't count on
your advisor to hand you all of the relevant papers
out of his file drawer. He doesn't have them all!

d. Thesis proposal

- It is the most crucial stage in the sense that the basic
concept is worked out here.

* To get important results you need to ask important
questions

* This is the time you need your advisor most.

* Problems in later stages are usually rooted from a
weak thesis proposal.

- Purpose

* A research plan

- A serious attempt to get an overview of the
whole research course

- Not really a contract

* Need some flexibility because research
always has uncertainty.

* Forming the committee

- Varies a lot

- Choose people for your thesis committee that
can help with needed expertise. For example,
it is useful to have a relevant theory faculty
member on a systems committee and vice-versa.

- However, there is usually no need to optimize
too much on the selection of the committee
members--advisor still plays the most important
role.

- However it can be very important, when

* you have a "questionable" advisor, or

* you have an interdisciplinary topic.

* A review

- If there is any serious doubt, it had better
show up now.

- Proposal could sometimes be viewed as just a
forcing function for taking care of certain
things.

- Some of the difficult questions always asked in a thesis
proposal:

* What is your approach and what is new?

* What is your secret weapon? (Herbert Simon)

* How do you measure your own progress?

* What are the success or completion criteria?

* How will the expected results change the-state-of-
the-art?

- The grand challenge for a thesis proposal is to come up
with an approach or an experiment.

* It is easy to identify a general problem area, but
setting up an approach and designing an experiment
can be difficult.

- Need ideas

* Just need one good idea, really

* Unfortunately, there is no magic here
(however see some hints below). This is
the hard part of any research project for
everyone (not just for students).

* Need independent thinking

- You should be good enough to start arguing with
your advisor on technical issues and research
tastes.

* Need to elaborate on focus, approach, experiment,
and potential impact

- For theory research you may propose some new
models of computation.

* Examples: area-time complexity (new VLSI
model in theory), parallel algorithms (new
cost models)

- For system research you may design experiments
and argue their relevance.

* Examples: multiprocessor architecture,
compiler for a parallel machine

- Useful things to know when preparing a thesis proposal

* Be honest. There is no need to exaggerate your
claims! If you point out the weaknesses in your
approach you will disarm your critics.

* Pick a project that is manageable so you can do an
excellent job - things are always harder than they
seem. It is far better to do an outstanding job on
a moderate size project than a moderate job on a
large project.

* Include a tentative thesis outline and a month by
month schedule in your thesis proposal.

- This may be difficult to do but it is better
than no plan at all.

- This will also help gauge the total size of the
work you are committing yourself to do.

e. Producing results

- Lots of work--what else do you expect?

* System--be inside an active project without losing
sight of thesis

- Need to be a worker as well as a conceptual
person.

- Your work depends on other people's work and
vice versa

* Opportunity to see real problems

* Getting good support, including
encouragement and demand, from the group

- It seems that this arrangement really
works in all cases.

- Be quick, because you don't want to be
overtaken by the environment (this is one of
the pitfalls to avoid, as described below)

* Theory--be lucky!

- Be flexible

* It is hard to insist that you will prove a
theorem before you go to sleep.

- Be quick, because theoretical results are
totally portable and so competition can be
keen.

- Keep the committee informed (at least those "trouble
makers")

* You can get real help sometimes.

* Committee members are obliged to talk to you.

- Sometimes finding a qualified person beyond
your advisor to discuss your work can be
difficult.

* Don't want surprises in the later stage of the
thesis

- Ways to finish a thesis

* Incremental and adaptive approach

- A sequence of incremental results

* Big-bang approach (this is not recommended in
general)

- One big theorem

- A big piece of software or hardware

f. Writing

- Why some students find that Ph.D. thesis writing is very
difficult

* First major document

* Writing is time-consuming--part of the .9999
perspiration (Satya)
- Think how many good sentences you can write in
an hour.

- Fighting with fonts, figures, references, etc.?

* Please don't be too picky.

* When results are not totally solid, writing can be
really difficult even for an experienced writer (now
you know another reason why proposal writing is not
easy)

- Can't say too much and don't want to say any
less

- Writing about flaky results can be a real
challenge.

* In this case you should improve your
results first.

* Writing has to do with presentation rather than
finding new results. So writing may not be as
exciting..

- However, thesis writing is useful in the sense that it
helps reveal possible problem areas and provides new
insights.

* Help get a large picture on what you really have.

* Help organize the concepts

* Completeness is forced.

- You must take care of things that you have been
ignoring.

* For example, you need to do comparison
with other results

* Correctness of the results is checked.

- You had better have the proof now for any
plausible "theorem" that you have been
believing.

* New insights on how things really work

- New ways of looking at you results

- Recommendations

* Get some practice--write some papers before thesis

- Write some joint papers with people who have
substantial writing experience

* Need to know the theme of the thesis very well

- Outline first

- Write the conclusion first (try it at least)

- Start writing chapters which are more settled.

- Write the introduction last

- Iterative process

* Make the writing as precise as possible, so that you
know exactly what you are talking about. This will
save lots of rewriting.

- Precise writing usually also yields good
English.

g. Getting final comments from the committee

- Not too early or too late

* Getting some committee members to read can be a
challenge.

- They are busy people. You want to give them an
"optimal" version to make comments.

- How much to ask for comments varies a lot

- Should not have any surprises now.

* You had better know what you have been doing by now.

* However, if there is any problem, it had better show
up now.

h. Defense

- Mostly a formality and a happy occasion (should be like
that)

* You know that your results are good and you will
present them well.

- You should know the answer to the question -
"What are the three main ideas in your
thesis?". You should be able to rattle them
off and relate them to previous work.

* Getting a date set can be more difficult than you
think.

- Committee members do not necessarily stay at
CMU as long as you do!

- Weekend defense is not really desirable.

* May be difficult to get audience.

- However defense is still very important:

* Opportunity for final improvements for the thesis

* Formal presentation to the community

- Many people form their opinion of your n-years'
work from this presentation

* Presentation material can be used for future
presentations

- Used in recruiting presentations if you have
not settled on a job yet

* Psychologically important

- Once in a life time occasion--you will remember
it always.

* Don't want to blow it.

- Absolutely no surprises

i. After defense

- Usually there is still some minor work to be done for the
thesis (too bad)

* Defense was moved early for various reasons

* New comments from defense

* Did not have time or did not want to polish the
thesis before defense

- Publication

* Articles, books (or give the thesis to your parents)

* Very important to publish the results in journals

- This is the only reliable way to archive your
results. (You don't want to lose them after
all these efforts, do you?)

- Publication is important for academic career.

- May break the thesis up in several articles.
When appropriate, some articles may have joint
authors such as your advisor.

- Do it right away before you get on to the next
thing.

* Books can be good too.

- Follow-on work

* Keep mining the thesis--why not?

- Finally you are free!

6. "Methods" to get into the depth of a topic (or how to come up with
good ideas)

- No magic, but we will still try ....

- How to develop initial ideas

* Study other work and do comparison

- What are similar issues and solutions?

* Look at examples

- Generalization and abstraction

* Make hypothesis and validate it formally or informally--
keep trying
- You will discover issues at least.

* Do modeling and abstracting

- Get the essence

* Just do something--be active

- Implementation--details reveal issues

* Join a project to do some real work!

* Handle a smaller case

* Implement a throw-away simulator, language,
design, etc.

- Start proving "theorems", even if they are known to
be difficult.

* Quick way to understand issues

* Work with good, experienced researchers (don't forget to
use your advisor!)

- They might have deep insights on similar problems.

- They can help calibrate the difficulty of the
problem.

- You learn the subject matter from them more quickly
and directly.

- You learn their techniques

* Every successful researcher has his or her own
bag of "tools":

- Calculation, synthesis, analysis,
persistence

- If they also get stuck once in a while, you know that
you are not that bad after all.

- How to develop existing ideas further

* Exploring problem and solution spaces

- Enumerate parameters individually (and do quick
pruning)

* To see where your current ideas sit in the space

- Correlate results

- Generalize ideas and results to other points in the
space

- Produce phenomena and explain them (Herb Simon)

* Brainstorming your ideas with others

* Presenting your ideas in papers or/and seminars

- Ideas will be checked out carefully and
systematically (see above on thesis writing)

* Example steps that can be used to get some depth from a
simple result such as a speed-up curve

- Explain the curve

- Look at the problem and solutions spaces

- Do some comparisons

- Change the assumptions

* How stable is the result?

* How will results vary or correlate under
different assumptions?

- Derive some general principle

* Similar curves for other situations?

- General comments

* Thinking is the key

- Thinking is more important than reading

* Books are not always right.

- Note that in the system area with few
exceptions people who build systems do not
have time nor need to write up their
experience--it is too bad but it is a
reality.

- Be alert on all sorts of opportunities

- Do the thinking right away while you have it.

* Ideas and interest may be lost more quickly than
you like to believe

* Talking to people

- Don't over do it (you still need to do the work
yourself)

7. Breaking myths

a. "Advisor is a stronger researcher than you."

- It is true that advisor is experienced, wise, smart
(maybe), and knowledgeable in general. Advisor also sees
a bigger picture, and has contacts in the area.

- However, advisor is not always right.

* Advisor is not as focussed as you.

* Advisor does not have more time or energy than you
do.

* Advisor is not as innovative in general.

- They know too much.

- They are more conservative.

* They know too many horror stories.

- Aging does not help.

* Advisor's knowledge may be obsolete (don't say this
in front of him or her!).

- You must believe that you can do better than advisor for
some research areas.

b. "System theses take longer than theory theses."

- The most difficult part of a thesis is to come up with
some good, new ideas. The difficulty in getting new
ideas is the same for theory or system research.

* Theory thesis is in general not about solving open
problems.

- Actually good theoreticians always work on new
problems, models and methods so that they can
solve the problems that are "solvable" in the
first place.

* Greatest contributions are ground breaking
ones, such as new models.

* New approaches give new insights to old
problems. This is the way open problems
usually get solved (e.g., the four-color
problem).

* For systems theses it is important that the major
ideas in the thesis are independent of the
implementation--the goal is to have the ideas live
on in other systems as well. A good systems thesis
usually has a new algorithm or new method at its
core.

* Few theory students who finish really early are
likely those who have prior research experience.
(Recall that theory results are highly portable!)

* Incompetent theory students are more noticeable than
weak system students. So we don't often see theory
students who drag on for a long time.

- There are some differences in systems and theory research
however, but they should not have too much impact on the
thesis research time.

* System needs implementation, whereas theory needs
more background study.

* Theory research is self-sufficient and system
implementation may depend on other people's work
(you should not get into a situation where you don't
have control).

c. "Ph.D. thesis research follows some standard guidelines."

- Yes, a Ph.D. this must represent a substantial result in
a very high standard.

- But there are many ways to leave a mark in a research
area. As long as you have come up with some good ideas
and pushed the frontier of knowledge, you will be
surprised sometimes how flexible your committee could be
in terms of the research approach, acceptable results,
and thesis presentation.

- There is a small percentage of Ph.D. theses completed in
unusual manner. Don't give up too early if you belong to
this class. Try it or you will never know.

8. Pitfalls to avoid (easy ones to avoid listed first)

a. The goal is too big to reach.

- Theory

* Proving P /= NP

* Proving P = NP is even worse (likely this thesis
will never finish!).

* Deciding whether P = or /= NP is best of the three
(i.e., be flexible)

- System

* The initial effort is so large that real issues
never get a chance to be looked at.

* It is important to size the project and evaluate the
total effort carefully based on past experiences.

b. Ideas cannot stand without an implementation that competes
with commercial products.

- Chess machine implementation is OK, because there is no
commercial competitor.

- In this sense, Warp hardware is more difficult than
software.

- Floating-point designs that require a high-performance
chip implementation to validate the concept would be
disastrous.

- Never need to implement another vector processor!

c. The thesis area is overtaken by technology and environment

- Technology advances have solved the thesis problem.

* A clever operating system using no more than 128K
memory is not very interesting today.

- Advisor (or student sometimes) has changed his or her
interest

- Other new projects have better approaches and
opportunities

- Other people have published similar and/or better
results.

- Advisor has a better job elsewhere or the project is
over.

- Lesson: You should always do your thesis as quickly as
possible.

d. Totally isolated work

- No encouragement and support--no one cares about your
thesis

* Can't even find an advisor sometimes

* Doing a thesis away from CMU is really difficult.

- System research

* Lone ranger approach is almost suicidal.

- No software, systems and application support
for evaluation

- Very difficult to do anything real without
feedback from a community

- Theory research

* At least global networking is needed.

e. Not knowing when to stop

- Thesis is not the last research you will do.

- You can do the same research after your Ph.D. thesis
(while making more money).

- Learn to make reasonable assumptions to restrict the
problem

f. Unhealthy competition between student and advisor

- This is more likely to happen in the theory area.

- The potential is always there (especially for smart
professors with lots of ego). In general if both sides
try to be fair, things can always be worked out.

g. Lots of numbers and hacking but no fundamental principles

- System research has to have more than implementation.

- Implementation for a thesis research is interesting only
if it can be used to validate some theory.

- This problem should be fixed as early as possible.

h. Things dragged on--wonderful general ideas in the beginning
that never get developed into a coherent approach (i.e.,
heading to a black hole--there is no output)

- Wrong areas for the student (and perhaps the advisor)
with respect to ability and interest

- Nightmare case--it does no good to anyone.

9. Some other general advice

- Stay away from areas that have been thoroughly mined by your
ancestors.

* Keep yourself at the very front of a research area so that
you have a better chance to hit something big or at least
new.

* After all in research what matters is the work that pushes
us into new territories.

* Make use new advances in other areas

- Don't avoid thinking

* Thinking is hard but there is no substitute for it.

- Psych yourself up for this unique experience of doing a Ph.D.
thesis

* Make yourself believe you are solving the most important
problem in the world

* Remember what worked for you before

- If you work best when you are competing with others,
then create some confrontation.

* Must be very alert about issues and opportunities

* Thesis process is sort of artificial (almost a torture in
some way)

- The thesis is judged by a committee (mainly your
advisor)

* More subjective than exams

- Probably one of the most humiliating experiences for
people of this age (advisors should all remember this
and be considerate.)

- The process is not a typical research style--you
don't do anything similar to it again even if you
will be doing research after the degree.

* The thesis process can be long and treacherous. (Be
prepared for it.)

- You don't want depression.

* There are quite a few very competent people who just do
not want to go through this.

- Use forcing functions well to speed up the thesis process

* Competing with someone else

* Family pressure

* Financial pressure

* A job is waiting

* Advisor is leaving or project is over

* Equipment is retiring

- Never throw away advisor's comments

* Cox-Denning case

- Keep good relationship with your advisor (even after you
graduate)

* Good thing to do--no exception almost

* Relationship is unique.

- Advisor usually has lots of influence on you in this
very important stage of your life. Advisor also
appreciates the good research you did with him, and
is in general interested in your well-being.

* Advisor may be your mentor for your entire career.

10. All the effort is worth it (believe it or not)

- Experience from Ph.D. thesis research is unique. You have
learned how to do research. Future research is going to be
more interesting because you will know how to do it, so you
will have more freedom and fun.

- Almost all leaders in research have this experience. You will
have confidence in your research ability. You will look at
things differently than people who did not go through this
process. It is very clear that Ph.D. thesis research is still
the best way we know of in developing powerful researchers.

- In summary, it is the best investment for becoming a successful
researcher.