Regression
Gradient Decent
- 根据loss function的斜率找到loss function的local minimum- learning rate决定步子大小
- 梯度正负决定移动方向(负 增加,正 减少)
- linear regression不需要担心saddle point,平缓处
-
Adaptive Learning Rate
- 不同参数不同learning rate- 随iteration渐渐缩小
Adagrad
Error
一般情况下,从Large bias & small variance => Small bias & Large variance
不同testing结果不一样,保证结果:
- Training Set 分两组:
- Training Set: train model
- Validation Set: 看小模型结果选mode
- 选完后用全部training data跑 最优model,再进行test
- 这样testing data结果较为真实, avoid public training data bias
Bias:
- unbias: 期望值等于u- 期望值estimator 与 靶心 中的距离 为bias
- Error 来自 bias过大:under-fitting
- Model cannot fit training data
--> Introduce more features OR choose a more complex model
Variance:
- 瞄准了靶心射偏了:variance- Error 来自 variance过大:over-fitting
- Model cannot fit testing data
--> Introduce more training data OR Regularization