Classification as Regression?
Class 1 and class 2 are more similar, class1 and class3 are not so similar.
1 Class as one-hot vector
2 Regression v.s. Classification
3 Soft-max
More commonly, it is always use sigmoid when binary classification (only 2 class), but sigmoid and soft-max are equivalent.
4 Loss of Classification
Minimizing cross-entropy is equivalent to maximizing likelihood.
In Pytorch, when you call cross-entropy, it will include soft-max.
Changing the loss function can change the difficulty of optimization.
5 Classification
5.1 Function with Unknow Parameters
5.2 Loss of a function
5.3 Training Examples
What do we want?
Probability of Failure
Tradeoff of Model Complexity -> Deep Learning
