In fact, it only took a little time to study the book today,
If the model has too many parameters, and the training data is not enough, there will be overfitting.
Overfitting can be solved by regularization, the Bayesian method can also avoid the appearance of overfitting, in fact, in the Bayesian model, the effective parameters of the model is automatically determined by the size of the training data set.
Regularization's idea is that adding penalty to the error function makes the coefficients not very large. In the Li Hongyi video, the problem is also spoken, but every one understands and understands.
The complexity of the model should be determined by the complexity of the problem to be solved, rather than the number of test data sets, which is more reasonable.
Persistence, patience, effort!
DAY3----"Pattern Recognition and machine learning" Christopher m. Bishop