Jethro's Braindump

Article: An Opinionated Guide to ML Research

tags
Machine Learning, Research
author
John Schulman
source
An Opinionated Guide to ML Research

The keys to success are *working on the right problems, making continual progress on them, and achieving continual personal growth.* This essay is comprised of three sections, each covering one of these topics.

Choosing Problems

Develop good taste for what problems to work on.

  1. Read a lot of papers, and assess them critically.
  2. Work in a research group with other people working on similar topics.
  3. Seek advice from experienced researchers on what to work on.
  4. Reflect on what research is useful;
    1. When is theory useful?
    2. What causes some ideas to get wide uptake?

Idea-Driven vs Goal-Driven Research

Idea driven
follow some sector of literature, and improve on existing ideas. - Cons: High risk of getting snooped, or duplicating work, requires deep understanding of subject
Goal-driven
Have vision of new AI capabilities, and experiment with diferent methods across the literature. Develop your own methods to improve on them. - Pros: much more motivating. goals also give differentiating perspective from rest of the community.

John Schulman recommends goal-driven research.

Tips for Goal-Driven research

  1. Don’t take the goal too literally. Restrict yourself to general solutions.
  2. Aim high, and climb incrementally towards high goals
    1. Keep a notebook, and record daily ideas and experiments (Note-taking). John Schulman creates an entry for each day, and conducts a review to condense entries every 1 or 2 weeks.
  3. Know when to switch problems. Also, don’t switch problems too often. Notice the dead-ends in half-finished projects, and if there aren’t any, make a commitment towards following through in the future.
  4. Have a fixed time budget for trying out new ideas that diverge from the main line of work.

Personal Development

  • Read textbooks, theses and papers, reimplement algorithms from these sources. Especially textbooks, because they condense information in an ordered fashion, and using proper notation!
  • Theses are a good place to find a literature review of an active field.

Links to this note