- Arthut Samuel: Field of study that gives commputers the ability to learn without being explicitly programmed.
We need to find relationship between X and Y.
In supervise learing, you are given a dataset with inputs X and labels Y. And our goal is to learn a mapping from X to Y.
The problem above is a regression problem. The value y you're trying to predict is continuous.
Tumor is belign or malicious.
In most cases, the imput x will be multi-dimensional.
Instead of just using tumor size to predict malignant vs. benign tumors, We may instead have two features. One is tumor size and the other is age of patient. Use the two features to predict a given tumor is malignant or benign.
Other features concerned about:
- clump thickness
- uniformity of cell size
- uniformity of cell shape
SVM uses infinite number of input features. to represent a patient.
Machine Learning Strategy
Towo teas will have definitely different effiency wth the sane learning algorithm.
We have many decisions to make during the research. We should be more systematic when driving machine leanring as a systematic engineering discipline.
An unsupervise learning will be no labels. We just get x and no Y. And we are asked to find something interesting in this data. Some algorithms like K-means cluster will discover this.
Cocktail Party Problem
We are in a noisy room and we stick multipe microphones in the room and get overlapping voices. How can we get an algorithm separating the noises.
The algorithm we will use is called ICA, independent Components Analysis.
Stanford autonomous helicopter
Just tell it good when it does the good thing and tell it bad when it do the bad thing. Because we do not know the optimal way to fly a helicoper, so we just let helicopter do whatever it can do and we tell it what is good and what is bad.