How to read deep learning research paper ?
Research Paper Reading Approach - A notes from multiple legends lecture eg.Andrew Ng and more .
How to Read the Research Paper.
Compile the List of papers,few places to look are below
Arxiv Sanity.
paperwithcode
research gate
DeepMind
Twitter
ML sub reddit
NeurIPS/ICML/....
GOAL
Read 50-100 Research paper to have very good understanding of that Domain.
How to Read One Paper
Multiple Passes.
1. First pass --- Title/abstract/Figures.
2. Intro -- Conclusion -- Figures -- skim rest.(skim the related work)
3. Read the paper but skip the Maths.
4. Real the whole thing but skip the part that doesn't make sense.
5. what did the authors tried to accomplish ?
6. what are the key Elements of the approach ?
7. what can you use yourself ?
8. what other references do you want to follow ?
Now in details ?
• Ask questions before you start
• What is the motivation?
• What is the problem?
• What is the solution?
• What experiments are done?
• How is the solution evaluated?
• What are the contributions?
• Future direction?
• Relation with existing works?
The Process?
• A one-time reading is not going to work
• Multiple iterations
Step 1:
• Read the title
• Scan all the sections
• Outcome
• It will give you an idea what the paper is about
Step 2:
• Read the abstract
• Scan all the sections again
• Look for figures about the method
• Read the conclusion
• Outcome
• It will give you an idea about
• The problem being solved
• A bit about the approach
Step 3:
• Read the abstract again
• Read the introduction
• Read the overview of the approach
• Try to understand the method from the main figure
• Look for the results which are shown
• Outcome
• Better understanding of the problem
• More insights into the solution
• Evaluation
Step 4:
• Read method very carefully
• You can skip alien words
• Try to understand as much as you can
• Go through the evaluation carefully
• Outcome
• You should have a better understanding of the paper now
• If you didn’t understand some part of the paper
• Now is the time to look into it