這講建構了一個大略的 ML 世界觀。
Different Output Space
- Binary classification ($|\mathcal{Y}| = 2$)
- Multiclass classification ($|\mathcal{Y}| = K$)
- Regression ($\mathcal{Y} \subset \mathbb{R}$)
Different Label
- Supervised (w/ explicit label)
- Unsupervised (e.g clustering)
- Semi-supervised (combine above 2, owing to label cost issue)
- Reinforcement (learn partial info. sequentially)
Different Protocol
- Batch learning (Given all known data once)
- Online learning (can improve when seeing new data)
- Active learning (interactively query label $y_n$ of chosen $x_n$)，也是很常見的蒐集 data 方式(欸！？)，跟 Online Learning 類似，但差別在會 主動 query
Different Input Space
- Concrete feature (利用human intelligence 抽取出有關連的特徵將 input 做簡化)
- Raw feature (通常需要再做 feature engineering，但現今 DL 的前半部通常會一起 end-to-end train 成一個還不錯的 feature extractor)
- Abstract feature (原先的 feature 沒有任何物理意義(e.g 可能只是 ID 而已)，需要做 feature conversion 轉成上述二者)