从上一讲我们得出结论：learning happens if finite $d_{VC}$, large $N$, and low $E_{in}$.这一讲主要介绍在有噪声情况下的学习问题以及相关损失函数。

1. Noise and Probabilistic Target

2. Error Measure

机器学习的终极目标是$g\approx f$,那么如何度量其相似度呢？

 1）Pointwise Error Measure

  Two Important Pointwise Error Measures

2）Learning Flow with Error Measure

3. Algorithmic Error Measure

  0/1 error penalizes both types equally.

 Learning Flow with Algorithmic Error Measure

4. Weighted Classification

 Weighted classification: different “weight”for different (x; y).

 Easily done by virtual "example copying"!

1) Minimizing Ein for Weighted Classification