search

bobye / neuron

Scala library for neural networks
109 stars 32 forks source link

Dropout training #20

Closed bobye closed 10 years ago

bobye commented 10 years ago

May have problem if a module has multiple instances. It requires to record dropout in the apply(), and later in back propagation can know which neuron to drop. Further it may be necessary to create another apply() for training purpose.

bobye commented 10 years ago

A possible way to work out this feature is to create a dropout layer (as a nonlinear transform)