Note 1 – The Learning Problem

Key Essence: Improving some performance measures with experience computed from data.
- Exits underlying patterns to be learned.
- No easy programmable definition.
- Data about the pattern.
Learning Flow
- Unknown target function $f : \mathcal{X}->y$
- Training examples: $\mathcal{D}: (x_1, y_1), ..., (x_n, y_n)$ .
- Hypothesis set: $\mathcal{H}$ .
- Learning algorithm: $\mathcal{A}$ .
- Final hypothesis $g \approx f$ .
Relative Fields
- Machine Learning vs Data Mining vs Artificial Intelligence vs Statistics
  - ML: Use data to compute hypothesis $g$ that approximates target $f$ .
  - DM: Use a huge amount of data to find a property that is interesting. Focus on efficient computation in large database.
  - AI: Compute something that shows intelligent behavior. ML is one possible way to implement AI.
  - Statistics: Use data to make inferences about an unknown process. Statistics can be used to achieve ML, focus on provable results with math assumptions.