An Opinionated Guide to ML Research

Abstraction: Practical advice on problem selection and research habits for ML researchers

Key points:

• Goal-driven research (pursue a specific capability) is preferred over idea-driven (follow and extend the literature) — reduces duplication risk and provides differentiating perspective that leads to unexplored ideas
• Author's own goal-driven arc: focus on 3D humanoid RL locomotion led to TRPO, GAE, and PPO — while others chased DQN/Atari bandwagon
• Keep a daily research notebook; weekly review with sections for experimental findings, insights, code progress, and next steps; also maintain a backburner ideas list
• Switching problems too frequently is a more common failure mode than insufficient switching — resist chasing new ideas before current thread yields a deliverable
• Allocate explicit time for textbooks and thesis reading: textbooks offer denser knowledge than papers; PhD theses contain unified perspectives not found in individual papers
• Reinimplementing papers from scratch yields much deeper understanding and faster feedback than passive reading

Connections: Openai · Deepmind · Reinforcement Learning · Machine Learning Research

Source: http://joschu.net/blog/opinionated-guide-ml-research.html