Learning with Different Output Space $\gamma$
Binary classification
Y = {−1, +1}
Multiclass Classification
- classify US coins (1c, 5c, 10c, 25c) by (size, mass)
- $\gamma = {1c,5c,10c,25c}$,or $\gamma = {1,2,··· ,K}$ (abstractly)
- binary classification: special case with K =2
Regression
- $\gamma = \mathbb{R}$ or $\gamma = [lower, upper] \subset \mathbb{R}$ (bounded regression)
Structured Learning: Sequence Tagging Problem
- a fancy but complicated learning problem
- sentence -> structure (class of each word)
- $\gamma = \{PVN,PVP,NVN,PV,···\}$, not including VVVVV
- huge multiclass classification problem ($structure \equiv hyperclass$) without explicit class definition
Learning with Different Data Label $y_n$
Supervised learning
- every $x_n$ comes with corresponding $y_n$
Unsupervised learning
- clustering
- articles -> topics
- consumer profiles -> consumer groups
- density estimation: {xn} -> density(x)
- i.e. traffic reports with location -> dangerous areas
- outlier detection: {xn} -> unusual(x)
- i.e. Internet logs -> intrusion alert
Semi-supervised learning
- leverage unlabeled data to avoid expensive labeling
Reinforcement Learning
- Teach Your Dog: Say Sit Down
- cannot easily show the dog that $y_n$ = sit when $x_n$ = sit down
- but can punish to say $\hat{y_n}$ = pee is wrong
- but can reward to say $\hat{y_n}$ = sit is good
- learn with partial/implicit information (often sequentially)
Learning with different Protocol $f \rightarrow (x_n,y_n)$
Batch Learning
- batch supervised multiclass classification: learn from all known data
Online Learning:
- hypothesis improves through receiving data instances sequentially
Active Learning: Learning by ‘Asking’
- improve hypothesis with fewer labels (hopefully) by asking questions strategically
Learning with different Input Space $\chi$
- concrete features: each dimension of $\chi \in \mathbb{R}$ represents sophisticated physical meaning
- Raw Features
- simple physical meaning; thus more difficult for ML than concrete features
- often need human or machines to convert to concrete ones
- Abstract Features: again need feature conversion/extraction/construction
