Translation of Andrew Un's book, Passion for Machine Learning, Chapters 55 - 57

previous chapters


55. The general case of attributing errors to a component of an integrated system


Below is a sequence of steps to correlate the errors of a complex system with its specific module. Consider a conveyor consisting of three components A , B and C , wherein the result of component A is transmitted to the input B , and output B at input C .


image


For each system error in the validation sample, do the following:


  1. To try to manually create a "perfect" output signal from component A (for example, the "ideal" image cat fragment), and run the modules B , C , submitting them to the input of the "ideal" output component A . If the system now gives the correct result, this means that if module A did its job more efficiently , the overall conclusion of the system would be correct; thus we assign this error component A . If the system continues to generate an error, then go to step 2.
  2. It is necessary to manually change the output of module B , making it “ideal”. If the algorithm now gives the correct conclusion, then we attribute the error to this module. Otherwise, go to step 3.
  3. C.

:
image


«». () ?


:


A:


B:


C:


. , , , , , . . :


  1. A ( ), «» (, , ). B, C, , C ( ) « », A. , , , A , ; , . ( ) 2.
  2. ( B ), . , , B. 3.
  3. C.

(DAG), , , () (). A -> B -> C, DAG, . , :


A: ( )


B: ( )


C:


, , .


56.


, , (data science) , , , . , , , , .


-. (data scientist) . . , , , (lifetime value) . «» . «» . , , .


, « » (, , ) , « » . « », .


image


, , , , :


  1. « » ?
  2. « » ?
  3. ? , ( ). , , « » , , «».

, , , , .


, , , . . , , .


, — , .


57.


, , , , ?


, . .


:
image


, . , . . .


  1. « » ()
  2. « » .
  3. « » , , «» , .

, , , . - , . ?


, , . « » , , . , , , «» , . , , ?


, , - . :


image


: , ( « »). , , 51. . « » — , , .


To summarize: if there is an understanding that a system consisting of modules, the quality of each of which is not inferior to human (you always need to remember that with such comparisons, we suggest that a person solve problems having the same information as the system modules), but Since the complex system does not reach much human performance, it is necessary to redesign the entire system.


final chapter


All Articles