Popular outlines for machine learning papers
Rob Zinkov
2015-06-05
When reading machine learning papers, I couldn’t help but notice that I kept seeing the same structures over and over again. Some conferences have one more often than the others. So, I decided to catalog a few of them.
The Theory Paper
This is the paper with the simple algorithm. You get this when you proved a nicer bound or the main contribution of your work is a faster convergence rate.
- Introduction
- Background (Graphical Models)
- Algorithm Details
- Theoretical Results
- Related Work
- Experiments
- Discussion
Note even if the algorithm is trivial, you will still see an experiments section. As near as I can tell, this is just because reviewers want to know you didn’t do this entire paper on a whiteboard.
The Systems Paper
This paper is for those projects that are more software engineering than new algorithms. Open-source libraries will have paper with this structure. Expect something in here about how contributors or lines of code the software.
- Introduction
- API
- More API
- Implementation
- Related Work
- Future Directions
- Conclusions
The primary thing this paper advertises is how easy it is to use a particular piece of software.
The Programming Languages Paper
This is really a systems paper when you see it show up in a machine learning conference. PL papers that end to be more theoretical will just go to POPL or ICFP. This comes up a lot for probabilistic programming.
- Introduction
- Language Description (types, terms, semantics, compilers)
- Example Programs
- Implementation Details (samplers, solvers)
- Discussion
The Algorithms Paper
This happens when you made a really complicated algorithm, that solves a particular problem. This comes up often because this algorithm is to solve a problem specific to a particular domain. You’ll see this for natural language processing and computer vision papers. Theory is something that is given little space in these papers.
- Introduction
- Algorithm Details
- Example Problem
- Experimental Results
- Related Work
- Conclusions
Note, this isn’t a systems paper. APIs are not talked about and you won’t be seeing many diagrams about the different components of the system. The authors had a concrete problem and this was how they solved it.
More to come
Think I missed a style? Tell me and I’ll add it.