Research tips
Collection of tips based on the following content:
- Guideline for Machine Learning PhD students by Markus Heinonen.
- PhD handbook by Stefano Albrecht.
- How to do good research by Eamonn Keogh.
- awesome-tips by jbhuang0604.
Research problems
A good research problem should be
- Important.
- Amenable to incremental progress rather than all-or-nothing style progress.
- Have a clear metric for knowing when you are making progress.
- Ideally applicable to real-world data.
Finding good problems is a hard task. Instead of finding solutions, focus on finding good problems, and understand them thoroughly. Try to focus on:
- Reading lots of papers to see what is missing, unaddressed and/or skimmed over.
- Find and understand latent needs. "If I had asked my customers what they wanted, they'd have said a faster horse. - Henry Ford".
- Extending an existing method:
- Make it more accurate.
- Make it faster.
- Explaining why it works so well.
- By applying it to a novel setting.
- Rethink some assumptions.
The research statement should be falsifiable. Examples of falsifiable statements include:
- Algorithm X is faster than Y.
- Function \(f\) is a lower bound on some other quantity \(g\).
- Method X generally outperforms Method Y under conditions Z.
Examples of unfalsifiable statements include:
- We can approximately cluster DNA with DFT. (Any random arrangement of DNA could be considered a "clustering").
- We present an alternative approach through Technique X to enhance the visualization. The experimental results demonstrate significant improvement of the visualizations. (The words "enhance" and "improvement" are subjective and vague. This could be made falsifiable by for e.g. mentioning "We improve the mean time to find a visualization by a factor of K" or "We enhanced the separability of information A,B as measured by metric M").