- Question - How do we increase speed to learn and long-term retention?
- Context - Quantified learning
- Answer - See below
- Primarily for personal learning or novel discovery? - novel discovery
- What is the delta? - Time to learn, quality of learning
Method
- Time: 1 month
- 2 phases
- Learn a concept: cycle through the materials until learned, then activate review phase.
- Review a concept: this is triggered upon completion of learn phase.
15 total tests:
For each combination, make 50-100 flashcards. See Quantified learning for data structures.
(3 Concept subjects x 2 Representation types x 2 Representation captureModalities) + (1 baseline experiment for each of the 3 Concept subjects) = 12 + 3 = 15 combinations
- Subject vs Subject vs Subject (Concept subject): (1) Language learning (src: anki decks), (2) natural language/trivia (src: wikipedia), (3) deep learning/math (src: textbook/blogs/code)
- Flashcard vs Memory palace (Representation type)
- Type conversation vs Verbal conversation (captureMethod is constant, captureModality is changed)
- 3 baseline experiments:
- Language learning: anki, vs my own system
- Natural language/wiki trivia: make my own flashcards from scratch, vs generate/chat
- Deep learning/math related: make my own flashcards from scratch, vs generate/chat
Subjects & their sources
- Language learning
- B1 Spanish - online Anki flashcards
- Natural language / trivia
- Wikipedia articles about philosophy and religion.
- Deep learning / math
- Stanford’s CS 229 coursereader and problem sets. Equations and bitesized problems to solve
Metrics
The application will need to track:
| Metric | How to measure |
|---|---|
| Time to learn a concept | Timer (ms) during learning phase |
| Correlation between a concept and a representation | cosine similarity of embedding |
| Average time to review a card | Timer (ms) during review phase |
| Lifetime time to review a card | Timer (ms) during review phase |
| How well I feel I know this topic | Synthetically generated test based on source material and flashcards |
Potential risks
- Comprehension of topic is based on synthetically generated test which may be biased towards the learning materials generated by the system.
- Variation of source materials within a single subject
- Potential confirmation bias towards what I’d expect to be a more efficient system may lead
- Whether material that is being remembered is appropriately weighted for what the learner really would like to learn
- This is tested with n of 1. Maybe 2 or 3.
Future experiments
- Different types of flashcards
- Different types of memory palaces
- Other formats
- Vivid stories as a way to learn
- Memorize examples of using a formula
- Prompt variability
- Feedback mechanisms
- Testing intervals
- More subjects
Results
This will be filled in over time.
| Experiment | Time to learn a concept (ms) | Correlation between concept and representation (float) | Average time to review a card | Total time to review a card | Topic comprehension |
|---|---|---|---|---|---|
| 14 | |||||
| 15 | |||||
| 16 | |||||
| 17 | |||||
| 18 | |||||
| 19 | |||||
| 20 | |||||
| 21 | |||||
| 22 | |||||
| 23 | |||||
| 24 | |||||
| 25 | |||||
| 26 | |||||
| 27 | |||||
| 28 |
Tags
- personal mnemonic:
- umbrella: Quantified learning
- type: #memory
- theme: #learn
- status: #in-progress