machine learning - How to find dynamically the depth of a network in Convolutional Neural Network -

i looking automatic way decide how many layers should apply network depends on data , computer configuration. searched in web, not find anything. maybe keywords or looking ways wrong.

do have idea?

the number of layers, or depth, of neural network 1 of hyperparameters.

this means quantity can not learned data, should choose before trying fit dataset. according bengio,

we define hyper- parameter learning algorithm variable set prior actual application of data, 1 not directly selected learning algo- rithm itself.

there 3 main approaches find out optimal value hyperparameter. first 2 explained in paper linked.

• manual search. using well-known black magic, researcher choose optimal value through try-and-error.
• automatic search. researcher relies on automated routine in order speed search.
• bayesian optimization.

more specifically, adding more layers deep neural network improve performance (reduce generalization error), number when overfits training data.

so, in practice, should train convnet with, say, 4 layers, try adding 1 hidden layer , train again, until see overfitting. of course, strong regularization techniques (such dropout) required.