10. Advice for applying machine learning

Advice for applying machine learning

Deciding what to try next

Debugging a learining algorithm:

Get more training set. (Sometimes doesn't actrully help)
Try smaller sets of features. ()
Try getting additional features
Try adding polynomial features
Try decreasing or increasing $\lambda$

Machine learning diagnostic: save your time.

Evaluating a hypothesis

Test your hypothesis whether overfitting or not:

Split your training into 2 part.one is Training set (70%)，another is Test set(30%).

Training/testing procedure for logistic regression

Learn parameter $\theta$ from training data
Compute test set error (liner/logistic regression)
Misclassification error (logistic regression)

Model selection and training/validation/test sets

when using $J(\Theta)$ to select which model to choose, overfitting may be happen and the result the model perform best isn't generlized.
$J_{test}(\theta^{(5)}$ is likely to be an optimistic estimate of generalization error.

Split the dataset into 3 pieces.

training set (60%)
cross validation set (20%)
testing set (20%)

Use the cross validation to select model.
Use the testing set to test the generlazation error.

Diagnosing bias vs. variance

underfitting overfitting

Bias (underfit): $J_{train}(\Theta)$ will be high; $J_{cv}(\Theta)\approx J_{train}(\Theta)$
Variance(overfit): $J_{trian}$ will be low; $J_{cv}(\Theta)\gg J_{train}(\Theta)$

Regularization and bias/variance

To find a good $\lambda$
Try 0 0.01 0.02 0.04 0.08 ...10.24
Get many $\Theta$ .
Use cross validation set to compute $J_{cv}(\Theta)$ , pick the minium of these $J_{cv}(\Theta)$

There are two figure in the two videos before, it's very useful to help to understand how the cross validation set helps to get best model and best $\lambda$

Learing curves

$J_{train}(\theta)$
$J_{cv}(\theta)$

error to m (training set size)

If a learing algorithe is suffering from high bies, getting more training data will not help much.

If a learning algorithm is suffering from high variance, getting more training data is likely to help .

Deciding what to try next (revisited)

bias: underfit
varaance: overfit

Get more training examples: fix high variance
Try smaller sets of features: fix high variance
Try getting additional features: fix high bias
Try adding polynomial features: fix high bias
Try decreasiong $\lambda$ : fix high bias
Try increasing $\lambda$ : fix high variance

?著作權歸作者所有,轉載或內容合作請聯(lián)系作者