机器学习入门理论基础|莫烦

Overfitting 过拟合

L1 regularization:

L2 regularization:

Dropout reularization:

activation function AF

linear/nonlinear

AF 函数：relu, sigmoid, tanh

AF要可微分 （在Back Propegation反向传递中把误差传递回去）

在层数较多时，AF选择要避免梯度爆炸/梯度消失

CNN: Relu

RNN: Relu or tanh

Classifier

feature对于分类的拟合度

避免对于分类无意义的信息

避免重复性的信息

避免复杂的信息

Feature nomalization

线性回归 linear regression

minmax normalization->(0,1)

std nomorlization->(mean=0, std=1)

Training data: 70%

Test data: 30%

误差Error

训练次数Epochs

Accuracy

R2 score

F1 score

过拟合

交叉验证

Autoencoder

Encoder

LSTM: long short term memory

Gradient explosion: 梯度爆炸

Full connection (FC):

Pooling

image->Convolution->max pooling->convtolution->max pooling->FC->FC->Classifier

Input layer

hidden layer

Output layer

Activation

Semi-supervised learning

Reinforcement learning

Generic algorithm