周末上了一节 Yaser 的网上公开课，教授的发音虽然有点奇怪，但是为人风趣，循循善诱，课程内容也是深入浅出，既有干货又不至于太过枯燥乏味。看完之后记了一点笔记，记录于此，希望自己能学完这套课程。(因为是英文课程，就直接记英文了，也算练习下英文）
The Learning Problem
The essence of machine learning (three components)
- A pattern exists
- We cannot pin it down mathematically
- We have data on it
Target function: f : X -> Y
Data: (x1 y1), (x2, y2), ... , (xn, yn) (historical records)
Hypothesis: g : X -> Y (final hypothesis is the trained model, g is close to f)
learning model = hypothesis set + learning algorithm
How does it work
Perceptron Learning Algorithm (PLA)
- The perception implements: h(X) = sign(w * x)
- Given the training data: (x1, y1), (x2, y2), ... , (xn, yn)
- pick a misclassified point: sign(W * Xn) != yn
- and update the weight vector: w <- w + xn * yn (yn = (-1, 1), so w <- w +/- xn)
Types of learning
Basic premise of learning
"using a set of observations to uncover an underlying process" -- broad premise that be applied in many disciplines
- Supervised learning: See above
- Unsupervised learning: Instead of (input, correct output), we get (input, ?). Cluster is one of them.
- Reinforcement learning: Instead of (input, correct output), we get (input, some output, grade for this output). It does not give you the output directly, but when you choose an output, it will tell you how well you're doing. It is interesting because it is mostly our own experience in learning. The most important application of reinforcement learning is in playing games. (like Alpha Go?)
A Learning puzzle
The target function is unknown, it could be anything.
We have a finite sample, it can be anything outside.
I'm new in machine learning and my English is not good enough. So please talk to me if there are any mistakes.