I want to learn [SUBFIELD] through implementation, progressing from basics to state-of-the-art.
Create a learning progression with these requirements:
1. START from the simplest possible version of the problem where an intuitive/naive approach will partially work
2. Each level should introduce EXACTLY ONE new challenge or failure mode that motivates learning a specific technique
3. For each level, provide:
- A specific dataset (with download link or code snippet)
- The architecture to implement
- What failure mode the previous level's approach will exhibit
- The new concept/technique that addresses it
- A paper or resource to read AFTER struggling with the problem
4. Progress should cover:
- Classic approaches (pre-2015)
- Modern deep learning approaches (2015-2020)
- Current state-of-the-art (transformers, foundation models where applicable)
5. Each level should be completable in 3-7 days of focused work
6. End with a summary table showing: Level | Dataset | Architecture | Key New Concept
The goal is NOT to copy-paste solutions, but to:
- Hit specific walls that make me understand WHY certain techniques exist
- Build intuition through struggle before learning the "right" answer
- Have each technique feel like a discovery rather than arbitrary knowledge