Regulariza*on

 
The  problem  of  
overfi6ng  
Machine  Learning  
Example:  Linear  regression  (housing  prices)  
Price  

Price  

Price  
Size   Size   Size  

Overfi&ng:  If  we  have  too  many  features,  the  learned  hypothesis  
may  fit  the  training  set  very  well  (                                                                                          ),  but  
fail  to  generalize  to  new  examples  (predict  prices  on  new  
Andrew  Ng  
 Example:  Logis*c  regression  

x2   x2   x2  

x1   x1   x1  

(        =  sigmoid  func*on)  

Andrew  Ng  
Addressing  overfi&ng:  
size  of  house  

Price  
no.  of  bedrooms  
no.  of  floors  
age  of  house  
average  income  in  neighborhood   Size  
kitchen  size  

Andrew  Ng  
Addressing  overfi&ng:  

Op*ons:  
1.  Reduce  number  of  features.  
―  Manually  select  which  features  to  keep.  
―  Model  selec*on  algorithm  (later  in  course).  
2.  Regulariza*on.  
―  Keep  all  the  features,  but  reduce  magnitude/values  of  
parameters        .  
―  Works  well  when  we  have  a  lot  of  features,  each  of  
which  contributes  a  bit  to  predic*ng        .  
 
 
Andrew  Ng  
 Regulariza*on  
Cost  func*on  

Machine  Learning  
Intui3on  

Price  
Price  

Size  of  house   Size  of  house  

Suppose  we  penalize  and  make          ,            really  small.  

Andrew  Ng  
Regulariza3on.  

Small  values  for  parameters    

Andrew  Ng  
Regulariza3on.  

Price  

Size  of  house  

Andrew  Ng  
In  regularized  linear  regression,  we  choose            to  minimize  

What  if            is  set  to  an  extremely  large  value  (perhaps  for  too  large  
for  our  problem,  say                                    )?  
Price  

Size  of  house  

Andrew  Ng  
 Regulariza*on  
Regularized  linear  
regression  
Machine  Learning  
Regularized  linear  regression  
Gradient  descent  
Repeat  

Andrew  Ng  
Normal  equa3on  

Andrew  Ng  
Non-­‐inver3bility    (op3onal/advanced).  
Suppose                                  ,  
       (#examples)    (#features)  

If                          ,  

Andrew  Ng  
 Regulariza*on  
Regularized  
logis*c  regression  
Machine  Learning  
Regularized  logis3c  regression.  

x2  

x1  
Cost  func*on:  

Andrew  Ng  
Gradient  descent  
Repeat  

Andrew  Ng  
Advanced  op3miza3on  
function [jVal, gradient] = costFunction(theta)
jVal = [ code  to  compute   ];

gradient(1) = [code  to  compute   ];

gradient(2) = [code  to  compute   ];

gradient(3) = [code  to  compute   ];

gradient(n+1) = [ code  to  compute   ];