1-【选修】逻辑回归
1 Logistic Regression
1.1 step1: Function Set
–
1.2 Step2: Goodness of a Function
cross entropy 衡量两个分布有多接近
1.3 Step3: Find the best function
1.4 Review
2 why not Logsitic Regression + Square Error?
3 Discriminative V.S. Generative
Logistics模型没有假设,但generative假设样本的probability distribution为高斯?朴素贝叶斯?等
3.1 which one is better
通常认为discriminative比generative要好
如上图所示,在Naive Bayes中并没有考虑不同dimension之间的correlation。所以generative模型中假定的probability distribution有可能会脑补出不应该有的条件
3.2 Benefit of generative model
4 Multi-class Classification
3 classes as example
5 Limitation of Logistic Regression
5.1 Feature Transformation
5.2 Cascading Logistic Regression Models
让机器自己学习找到好的feature transformation
这样机器自己学习后进行了feature transformation,从
x
1
,
x
2
x_1, x_2
x1,x2转到
x
1
′
,
x
2
′
x'_1, x'_2
x1′,x2′,再通过转化后的feature进行分类
Neural Network就来咯