← previous chapters

51. Choosing the components of the pipeline: ease of solving problems

The second important factor to consider when building a pipeline, in addition to the availability of training data, is the ease of use of the component. You need to try to choose such components of the pipeline, each of which is easy to develop or to learn. But what does it mean: the component is easy to learn?

The following are machine learning tasks in increasing order of difficulty:

1. Classification of overexposed images (as in the example above)
2. Determining whether an image is taken indoors or outdoors
3. Determining whether an image contains a cat or not
4. Determining if an image contains a cat with black and white fur
5. Determining whether an image contains a Siamese cat (a special breed of cat).

Each of them is a binary classification task: an image is input, the algorithm outputs either 0 or 1. Moreover, the higher the task in the list, the “easier” the neural network is to learn how to solve it. Simpler tasks require less training examples.

There is currently no good formal definition of task difficulty in machine learning

: « », , — , , ( ). . .

, «», ( (shallow neural network)), «», (, () ). .

, , , .

, . :

, :

( ) .

( ) , 1, - .

( ) , , 0 1.

: , , , . , , . , Deep Learning Coursera.

.

: (1) , (2) (3) . «», , .

, , , , «» , .

52.

, x , . , ( , , )?

x =

y = « , »

h: -> Y, y .

:

, , , y, , . , (x) (y).

The growing trend of deep learning is the teaching of end-to-end systems, the output of which, with the availability of suitable labeled data (labeled input and output for the system), you can get images, text fragments, audio clips and other complex objects, and not just numbers.

continuation