1 of 47

How to write papers & �How to give talks

Vincent Sitzmann�With thoughts from Vincent, Phillip Isola, and Bill Freeman

2 of 47

Announcements

  • Did you tag your team members when you submitted report on grade scope?
    • Everyone, please check if you got a zero b/c of this - no big deal, but have your team member tag you right now on the submitted assignment!
  • Grading rubric:�https://canvas.mit.edu/files/4208910/download?download_frd=1

3 of 47

Why to do research?

How

4 of 47

Why to do research?

  • Follow your own interest.
  • Be creative!
  • Learn the scientific method - how to find truth through hypothesizing, predicting, experimentation.
  • Discover something nobody knew before you.
  • Safe space to “be bold”, i.e., tackle big problems.
  • Fundamentally entrepreneurial: High “risk", high reward.

5 of 47

How to recognize good research?

6 of 47

Novelty

Good metrics: surprisal and enabling new directions.

What makes something surprising?

  • People should gain a lot of information from it
    • It should be new
    • It should be understandable

Enabling new directions:

  • Does this paper enable a new approach to an old problem?
  • Does this paper establish a bridge between two problems, such that tools from one are now useful for the other?

7 of 47

Novelty

Very hard to achieve without knowing what has already been done.

http://matt.might.net/articles/phd-school-in-pictures/

8 of 47

[Picasso]

9 of 47

[Picasso]

10 of 47

How to do good research?

11 of 47

A paper’s impact on your career

Paper quality

Effect on your career

nothing

Lots of impact

Creative, original and good.

12 of 47

Our image of the research community

  • Scholars, plenty of time on their hands, pouring over your manuscript.

13 of 47

The reality:

more like a large, crowded marketplace

http://ducksflytogether.wordpress.com/2008/08/02/looking-back-khan-el-khalili/

14 of 47

Picking a topic

15 of 47

Science is the “Art of the Soluble”

“‘Good scientists study the most important problems they think they can solve. It is, after all, their professional business to solve problems not to grapple with them.’ —Peter Medawar”

— Jitendra Malik

16 of 47

Know something no one else knows

  • It’s necessary but not sufficient to master the core knowledge of your field
  • Acquire a unique skillset
    • Take difficult or unusual classes
    • Read old papers
    • Take on a complementary hobby
    • Talk to people in other fields

“My answer to "Now What" is "here is a research problem which is unusual, perhaps significant, novel, that I can pose and probably solve because of my background in physics". The situation would not be readily identified as a problem at all by those whose background seems much more relevant than my own.” — “Now What”, John Hopfield

17 of 47

Question obvious weirdness

  • Often, when you are new in a field, you are confronted with something that everybody accepts, but that strikes you as weird.
  • These are often good research topics.
  • Example: Structure-from-Motion…

18 of 47

Build a ramp

Colorizing photos

Generative models

Agents that can imagine and plan

“When you are famous it is hard to work on small problems. This is what did Shannon in. After information theory, what do you do for an encore? The great scientists often make this error. They fail to continue to plant the little acorns from which the mighty oak trees grow.”

— Richard Hamming, “You and Your Research”

19 of 47

Think about the consequences

Why are you working on the problem you are working on?

What would happen if you were successful? Is that what you want to happen? Consider impact on science. Consider impact on society.

20 of 47

Get comfortable being confused

21 of 47

Omit needless bits

“Perfection is finally attained not when there is no longer anything to add, but when there is no longer anything to take away”

— Antoine de Saint Exupéry

22 of 47

Underfitting

K = 1

Appropriate model

K = 2

Overfitting

K = 10

23 of 47

Information creep

Number of terms in your loss function

Performance

24 of 47

Information creep

Number of terms in your loss function

Cognitive cost

25 of 47

Information creep

Number of terms in your loss function

Coding cost

26 of 47

Information creep

Number of terms in your loss function

Number of possible bugs

27 of 47

Information creep

Number of terms in your loss function

Generality of formulation

28 of 47

Information creep

Number of lines in GitHub readme

Chance someone will know how to use your code

100%

0%

29 of 47

Information creep

Number of experiments in your paper

Chance someone will understand your discovery

100%

0%

30 of 47

Information creep

Bits of information

Utility

Needless bits

31 of 47

Do the most, with the least

Discoveries

Explanations

Results

Tools

…

Words

Equations

Concepts

Lines of code

GPUs

People, Time, Money

…

Products

Costs

=

Metric for research

32 of 47

Pareto front of simplicity vs novelty

Simplicity

Novelty

Jibberish

Trivial

Simple and elegant

Complicated SOTA system

(useful information)

(1 / total information)

33 of 47

What the reviewing system rewards

Simplicity

Novelty

Jibberish

Trivial

Simple and elegant

Complicated SOTA system

(useful information)

(1 / total information)

34 of 47

What stands the test of time

Simplicity

Novelty

Jibberish

Trivial

Simple and elegant

Complicated SOTA system

(useful information)

(1 / total information)

35 of 47

The (abridged) Heilmeier Catechism

  • Problem definition: What are you trying to do?
  • How is it done today, and what are the limits of current practice?
  • What is your approach and why is it better?
  • Who cares? If you are successful, what difference will it make?
  • What are the mid-term and final “exams” to check for success?

George H. Heilmeier, former DARPA director (1975-1977)

36 of 47

How to communicate �your research?�(How to give talks)

37 of 47

High order bit: prepare

  • Practice by yourself.
  • Give practice versions to your friends.
  • Think through your talk.
  • You can write out verbatim what you want to say in the difficult parts.
  • Ahead of time, visit where you’ll be giving the talk and identify any issues that may come up.
  • Preparation is a great cure for nervousness.

38 of 47

Giving good talks is essential �to being a good scientist

  • You might think: “the work itself is what really counts. Giving the talk is secondary”.
  • But: you want to inspire people to help you work on the thing that you believe is the right way to go. By giving good talks, you can change the course of science!
  • Talks are a great opportunity to offer another angle on work that helps people understand it.
  • Researchers are like little startups: You have a brand and a product (your research). Talks is like doing sales: You want your product to help as many people as possible!

39 of 47

Your audience

  • Your image of your audience:
    • Paying attention, listening to every word

​

  • Your audience in reality:
    • Tired, hungry, not wanting to sit through yet another talk at the conference…

​

Your audience wants to be entertained.

Slide Credit: Bill Freeman

40 of 47

A tip to not be nervous that I found useful

  • Get over it. They’re not there to see you, they’re there to hear the information. Just convey the information to them.
  • Treat it like an opportunity for people to get to know you and what you’re excited about. Show them what you’re excited about, and they’ll like it :)

40

Slide Credit: Bill Freeman

41 of 47

A common mistake: Using your “talk voice”…

  • Somehow, we are taught that giving talks is somehow a different thing than explaining something to someone. That leads to people completely changing their body language and tone of voice.
  • Don’t. Giving a talk is essentially the same as explaining something to someone in-person, just with better preparation of the arguments and expositions :)
  • Use the same body language and tone of voice you would use in-person!

42 of 47

Body Language

  • If you can, walk around. Use the space. The podium is a trap!!
  • Use your hands.
  • Don’t look at the ground or computer screen all the time. If you need to look at the text on the slide, walk to the projector and point it out to the audience as you read it out.

43 of 47

Add dynamics to the talk

  • A talk is a story: there can be different levels of excitement or tension in different parts of the talk. This makes it easier for the audience to pay attention to what you’re saying. Perhaps move to another location.
  • I like to find some part of the work that really grabs me, that I’m really excited about, and let that show through. (The audience loves to see you be excited. Not all the time, but when appropriate).
  • People love hearing about when something confuses you - it makes you relatable. Paradoxically, stating that you were confused about something is not perceived as a “weakness” - it’s perceived as confident!

http://www.nch.ie/dynamic/img/Mariss%20Jansons%20%20new.jpg

http://images.google.com/imgres?imgurl=http://operachic.typepad.com/photos/uncategorized/ulster01.jpg&imgrefurl=http://operachic.typepad.com/opera_chic/2007/01/index.html&h=321&w=490&sz=57&hl=en&start=2&tbnid=6seUYhUX6x2DrM:&tbnh=85&tbnw=130&prev=/images%3Fq%3Dconductor%2Borchestra%2Bquiet%26gbv%3D2%26svnum%3D10%26hl%3Den%26safe%3Doff%26sa%3DG

44 of 47

Figure out how one part follows from another

Ahead of time, think through how each part motivates the next, and point that out during the talk. If one part doesn’t motivate the next, consider re-ordering the talk until it has that feel.

Slide Credit: Bill Freeman

45 of 47

45

46 of 47

What the audience of a technical talk wants

To have everything follow and make sense

To learn something

To connect with the speaker, to share their excitement.

They want to watch you love something!

​

Alan Alda: https://www.youtube.com/watch?v=j4XgjkXDxss, and others

46

Slide Credit: Bill Freeman

47 of 47

Let the audience see your personality.

  • They want to see you enjoy yourself.
  • They want to see what you love about the science.
  • People really respond to the human parts of a talk. Those parts help the audience with their difficult task of listening to an hour-long talk on a technical subject. What was easy, what was fun, what was hard about the work?
  • Don’t be afraid to be yourself and to be quirky.

http://is3.okcupid.com/users/112/250/11225140098321842389/mt1112532356.jpg

Slide Credit: Bill Freeman