[Machine learning course notes] Learning Map


[Notes of Machine Learning Course of Dr. Lee in NTU]

Purpose: Help me deeply understand Machine Learning.

Figures: Have not used the original figures of the PPT in the course. Redraw these following figures.

Difficult point: Structured Learning / Transfer Learning

#Machine learning ≈ Penetrating Tests

Define several inputs and observe the outputs. Difference between them is that whether the function is designed by the subject.

Penetrating Test

Machine Learning

The above figure of Machine Learning will be used following part of the article in several forms to visualize the models.


# The type of pairs of input & output.

  • Supervised Learning

Training Data contains input & output pair of target function.

Supervised Learning (Recognize cats and dogs)

  • Semi-supervised Learning

Unlabeled data as input but no standard for output. However, this scenario will actually improve the accuracy of the results resolved by the function.

Semi-Supervised Learning (Recognize cats and dogs)

  • Transfer Learning

Training Data is not related to the task considered, in other words, it can be either labeled or unlabeled.

Transfer Learning (Recognize cats and dogs)

#express thanks to l1nk3r

  • Unsupervised Learning

Machine need to generate output without input. It means it will be fed several random data without any identification or other attributes and it is asked to produce something meaningful.

e.g. Machine learns how to reading, drawing, etc.

  • Reinforcement Learning

e.g. Alpha Go

DeepMind (Source from DeepMind)

Q: What is the difference between supervised learning and reinforcement learning?

A: Supervised learning is a process machine learns from a teacher (selected Traning Data).

Reinforcement learning is a process machine learns from critics.

Q: What is the difference between Alpha Go and example in the Non-linear model in the following article?

A: Playing Go is a simple example of supervised learning. And it can only predict the next position in an inflexible resort.

Nonetheless, Alpha Go is a blended example of supervised learning and reinforcement learning.

Each Scenario has several TASKS

# The characteristic of output.

  • Regression


e.g. PM2.5 prediction; Temperature prediction

  • Classification

○ Binary Classification

Binary Classification

e.g. Spam filtering; 1vs1 winner prediction; Coin flips prediction

○ Multi-class Classification

Multi-class Classification

e.g. Document classification; Specified parameter vulnerabilities prediction

  • Structured Learning

This task is beyond classification, it cannot be simply determined as choosing from several samples. Actually, its input is utterly sophisticated.

e.g. Speech Recognition; Machine translation; Face Recognition

Each task can be divided into several MODELS

  • Linear Model

Non-linear Model

○ Deep Learning

Image Recognition (Each species is a class)

Image Recognition

Playing Go (Each position is a class, that is 19*19 classes)

Playing Go


○ Decision Tree

○ K-NN

○ ...


Machine Learning and having it deep and structured (2016, Spring), Hung-yi Lee.