Questions on lecture 12

[ladybird2012]

Hi,
I have 2 questions from lecture 12.

1) Slide 21: The graph on the RHS shows that when Qf=15 we need no regularizer. However, if I understand it right, this graph is based on the experiment performed on slide 13 of Lecture 11. On that slide we had overfitting when Qf>=10 since we were trying to fit the target with a tenth-order polynomial. So I would have assumed that for Qf>=10 we would need regularizer.... What am I missing?

2)Weight decay versus weight elimination for neural networks: I feel like these two regularizers are doing opposite things. Weight decay reduces the weights and favors small weights, but weight elimination favors bigger weights and eliminates small weights. So I guess these two regularizers are used under different conditions in neural networks -- could someone give me an example so I can pin it down? Are they ever both used in the same learning problem?

Thanks a lot in advance.

[yaser]

Originally Posted by ladybird2012

1) Slide 21: The graph on the RHS shows that when Qf=15 we need no regularizer. However, if I understand it right, this graph is based on the experiment performed on slide 13 of Lecture 11. On that slide we had overfitting when Qf>=10 since we were trying to fit the target with a tenth-order polynomial. So I would have assumed that for Qf>=10 we would need regularizer.... What am I missing?

The figure in slide 21 uses different parameters compared to the overfitting figures. The model being reguarized is 15th order, and there is zero stochastic noise in that part.

2)Weight decay versus weight elimination for neural networks: I feel like these two regularizers are doing opposite things. Weight decay reduces the weights and favors small weights, but weight elimination favors bigger weights and eliminates small weights.

Weight elimination does not favor bigger weights. It tries to reduce all weights, but it has a bigger incentive to reduce small weights than to reduce big weights.

[ladybird2012]

Thanks for the quick reply