Regression and Classification
-
binary (RNN)
-
multi-class (CNN)
generation
产生有结构的复杂的东西
supervised learning
-
labeled data
-
函数的loss, 预测值与实际值之间的关系,越小越好
-
机器会自动找出loss最低的函数
reinforcement learning
-
supervised: next move
-
reinforcement (reward): first moves …>many moves…> result
alpha go is surpervised +reinforcement learning
unsupervised learning
- no labeled data
network architecture
定义一个函数寻找的范围
函数寻找方法:gradient descent
meta learning
- learning algorithm
Regression
step1:model
y = b + ∑ w i x i y=b+\sum{w_ix_i} y=b+∑wixi
a set of function, input, output
step2: Goodness of Function
Define Loss function L:
-
input: a function
-
output: how bad it is
L ( f ) = L ( w , b ) L(f)=L(w,b) L(f)=L(w,b)
The effects of the loss function is determined by the selection of weight(w ) and the intercept(b).
In this pokeman example, the loss function can be defined as the sum of the differences between the predicted value and the actual value for each group of samples which is called as Least square method.
L ( f ) = ∑ n = 1 10 ( y n ^ − ( b + w x c p n ) ) 2 L(f)=\sum_{n=1}^{10}(\hat{y^n}-(b+wx_{cp}^n))^2 L(f)=n=1∑10(yn^−(b+wxcpn))2
- y n ^ \hat{y^n} yn^ is actual value in the train set
- b + w x c p n b+wx_{cp}^n b+wxcpn is the predicted value predicted by selected function model(linear model).
step3 Best function
pick the best function
f ∗ = a