AI, ML, and DL

Learning map

通常要用哪一個scenario是無法決定的，就沒資料就沒辦法做supervised learning

Regression

Applications

• Stock Market forecast

• Self-driving car

• Recommendation

Supervised learning

• Regression: The output of the target function f is "scaler," e.g., PM2.5 prediction.

• Classification: binary (spam mail or not) or multi-class (classifying news categories)

  • Linear model

  • Non-Linear model

    • Deep learning

    • SVM decision tree, K-NN

Semi-supervised learning

Very few labelled data with unlabeled data

Transfer learning

從別的有label的資料學到經驗，比方說，能夠判斷酒類的味覺和嗅覺特徵，那能不能拿來判斷美食的嗅覺與味覺特徵。

Unsupervised learning

e.g., Generating new paintings from a painting set

• Neighbor Encoding (Video, pdf): LLE(Locally Linear Embedding), Laplacian Eigenmaps (graph-based), T-distributed stochastic Embedding (t-SNE)

Structured learning

e.g., Machine translation, speech recognition, ...輸出的是一個有結構的東西。

Reinforcement learning

e.g., Alpha go 沒有告訴機器正確的答案是什麼，機器只能評估知道自己做的好不好，比方說電話交談，或者是整盤棋最後輸掉了。