The golden rule: do the most, with the least
(1-4 and 11 adapted from George Orwell's six rules)
- Rule 1: Never use a metaphor, simile, or other figure of speech that you are used to seeing in print
- Rule 2: Never use jargon if you can think of an everyday English equivalent
- Rule 3: Never use a long word where a short one will do
- Rule 4: If it is possible to cut a word out, always cut it out
- Rule 5: Remove any sentence that does not move the narrative forward
- Rule 6: One example is better than two
- Rule 7: Start with the concrete and specific, only then go general and abstract
- Rule 8: Vary sentence length
- Rule 9: If you see an adjective or adverb repeated, change it
- Rule 10: Never modify a noun with more than one or two qualifiers
- Rule 11: Break any of these rules sooner than saying anything outright barbarous
- Paper writing consists of four equally important parts: title, abstract/intro, figures, the rest
- A typical paper can only support one story; if you have two, kill one
- Figures should communicate the full story of the paper, without captions
- Kill your darlings
- Use le mot juste
- Develop your own voice
We will demonstrate by example:
- we start
with what is
the first... --> we startwith
the first... - there are many ways
by which one may
synthesize.. --> there are many waysone may
synthesize... although it may be that
policy gradient methods can... -->although
policy gradient methods can...Learning Visual Representations that Correlate with Auditory Stimulation
-->Learning to See by Listening
We note a few concurrent papers that
explore... -->A few concurrent papers
explore...The proposed work aims at developing
... -->Our goal is to develop
...- this work
highlights and addresses
... --> this workhighlights
... / this workaddresses
... For example,
infants... --> Infants...In this paper
we study --> We study
Simple but effective
,simple yet powerful
Novel
,new
Besides
- Repetitive structure:
We XX. We YY. We ZZ...
If you ever see the following, just cut the sentence completely and start with what comes after:
Humans have the remarkable ability to XX
,XX is a grand challenge in computer vision
, etcThere is an extensive literature on XX
,A great deal of work has studied XX
, etc
"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
We can think of any given project as having two stages. The first stage is an expansion of knowledge. We add components to our system to make it better and better, run a ton of experiments, and generally produce a lot of new information. The second stage, often overlooked, is the reduction of knowledge. In this phase, we take our system and remove any unnecessary component. If a system has five loss terms and four will get you within 1% of the performance, then just keep the four. Remove any experiments that are not crucial to proving your main points (or relegate these experiments to an appendix). The goal of this stage is to distill our knowledge down to its essence, without sacrificing anything truly important. Science is the iteration of these two stages, over and over.
Writing is the representation of an idea. Good writing is a good representation. What makes a good representation? Lucky for us, this is precisely the topic we study! Good representations are compact but informative; good writing is parsimonious. Good representations are disentangled; good writing deals with one topic at a time. Good representations are interpretable; you get the idea.
Most people are better critics than they are generators. If a colleague suggests you change a sentence, trust that the sentence really needs to be changed, but realize that your colleague might not know how to change it. Often we cannot express what is wrong with a sentence, we just know it is not right. Even if a colleague is able to provide a good rewrite, it dilutes your voice. An effective strategy is to keep rewriting your own sentences until all critics, including yourself, have no complaints.