Home
Home
Posts List
  1. Model Capacity and Overfitting
    1. Regularization
    2. Early Stopping
    3. Data Augmentation
    4. Dropout
    5. Hyperparameter Optimization

[Theory] [Deep Learning] ch6

Model Capacity and Overfitting

Regularization

add penalty terms P(W)
$$
L_r(\theta)=L(\theta)+P(W)
$$

  • L2: prefer smaller weight
    $$
    P(W) = ||vector(W)||_2^2
    $$

  • L1:
    $$
    P(W) = ||vector(W)||_1
    $$

Early Stopping

stop when testing error begins to increase

Data Augmentation

easy for classification, modify input sample a bit without changing class label

Dropout

During each minibatch, randomly set the output of some neurons in layer l to zero.

Hyperparameter Optimization

Grid search, try and error. Require a validation set.