notes

Personal notes
git clone git://git.laack.co/notes.git
Log | Files | Refs

Dropout.md (638B)


      1 # Dropout
      2 
      3 ML P604
      4 
      5 **Definition:** Dropout is a regularization technique for deep neural networks where upon each pass every neuron has a constant probability of being 'dropped out' meaning the output is 0.
      6 
      7 This works very well with a rate somewhere between 10%-50%. With RNNs we often do 20%-30% and with CNNs we use 40%-50%.
      8 
      9 This form of regularization ensures the model has multiple neurons that perform similar functions instead of being dependent upon one or only a few neurons to do important things.
     10 
     11 When using dropout we never drop output neurons. Additionally, it is common to only dropout neurons from the first few layers.