Research Advice

standford deep learning - Lec 8 career advice, Andrew Ng

• 5 ~ 10: basic idea
• 50 ~ 100 papers: good knowledge about the domain

Read paper through mutliple passes (not from first to last word)

1. Title, abstract, figure, experiment
2. Intro, conclusion, figure, skim (skip related work because it may used to acknowledge more authors to get paper accepted)
3. Read paper skip math
4. Read whole skip parts does not make sense yet e.g. in le net, some not irrevent anymore

Questions to ask when read paper:

• What do author try to accomplish?
• Key elements of approach?
• what can you use yourself
• Other references to follow

Sources of papers

• Twitter
• ML subreddit
• NIPS / ICML / ICIR
• Friends
• arxiv-sanity

Understand Math

• Rederive from scratch (e.g. art student copy from master then recreate)

2 ~ 3 papers per week instead of cram everything in a few days.

Karpathy

http://karpathy.github.io/

• Research Survive Guide

Prob not do incremental work e.g modify, combine, etc. Do interesting work, work that after e.g. graduate , is still valuable / add values.