How I Read Machine Learning Papers as a Freshman (Sample)
A practical routine for getting something out of a paper when most of the notation is still unfamiliar.
- learning
- machine-learning
- research
Sample content. This entry is placeholder material included to demonstrate the layout, not a record of real work.
Reading a paper front to back and understanding nothing is a waste of an afternoon. Reading the same paper in three passes usually is not.
Pass one: what problem, what result
Title, abstract, figures, conclusion. Nothing else. The goal is to be able to say what the paper claims and roughly how it argues for it. If I cannot, the paper is either above my level right now or not worth the time yet.
Pass two: the mechanism
Introduction and method, skipping every derivation. I want the shape of the idea: what goes in, what comes out, what is different from what came before. Notation I do not recognise goes on a list rather than stopping the read.
Pass three: the parts that matter
Only now the maths, and only the parts the result actually depends on. Usually that is one or two equations, not the whole section.
What I write down
For every paper, four lines:
| Field | Purpose |
|---|---|
| Problem | What was broken before |
| Idea | The one sentence version |
| Evidence | What convinced them it worked |
| Question | What I did not follow |
The last line is the useful one. It becomes the reading list.
The part nobody says out loud
Most papers are hard to read because they are written for people who already read twenty related papers. Not understanding one is information about the prerequisite chain, not about you.