Week03. Types of Learning

Course_Lin_MLF 2017.11

1. Learning with Different Output Space $\mathcal{Y}$

• binary classification
  实际中很多二分类问题，核心且基础

• Multicalss Classification：Coin Recognition Problem
  美国铜板分类，多种铜板
  $\mathcal{Y}=\{1,2, \dots, K \}$
  binary : $K=2$

• Regression:Patient Recovery Prediction Problem
  $\mathcal{Y} = \mathbb{R}$ or $\mathcal{Y} = [lower, upper] \in \mathbb{R}$(bounded regression)
  回归分析
  应用：股价预测，天气预测等等
  特点：输出是一个实数。

• Structured Learning：Sequence Tagging Problem
  自然语言处理中的词性标注，e.g. I(pronoun) love(verb) ML(noun).
  sentence $\to$ structure(class of each word)
  $\mathcal{Y} = \{PVN,PVP,NVN, \dots \}$，不包含$VVVVV$
  变成一个超大的多分类问题，structure $\equiv$ hyperclass，类并没有明确的定义

2. Learning with Different Data Label $y_n$

• Surpervised: Coin Recognition Revisited
  训练样本有对应的标签

• Unsupervised: Learning without $y_n$
  e.g.　Coin Recognition without $y_n$
  无监督，没有标签

  • clustering
  • density estimation
  • outlier detection

• Semi-supervised
  e.g. coin recognition with some label
  半监督的学习

• Reinforcement Learning
  增强学习
  e.g.　训练狗，对错误行为给予处罚，正确行为给予奖励
  implicit $y_n$ by goodness( $\tilde{y}_n$ )

3. Learning with Different Protocol $\to$ $(x_n,y_n)$

• Batch Learning
  一次给系统输入多个数据(一批，一个batch)进行训练

• Online Learning
  一个样本接一个样本依次学习更新

• Active Learning:Learning by 'Asking'
  Larning Philosophy
  batch: 'duck feeding'
  online: 'passive sequential'
  active: 'question asking'

4. Learning with Different Input Space $\mathcal{X}$

• concrete feature
  each dimension of $\mathcal{X} \subseteq \mathbb{R}^d$epresents 'sophisticated physical meaning'
  the easy 'ones' for ML

• Raw Features: Digit Recoginition Problem
  image pixels,speech signal

• Abstract Features: Rating Prediction Problem
  'no physicl meaning';thus even more difficult for ML
  abstract: again nedd 'feature conversion/extraction/construction'

• and more...