October 2018
Intermediate to advanced
246 pages
6h 26m
English
The dropout layer literally refers to dropping a few units of data by ignoring them randomly. This means that contributions to the downstream neurons are removed on the forward pass and the weights are not applied on the backward pass. If the neurons are missing during training, the other neurons will try to make predictions for the missing one. In this manner, neurons will become less effective for specific weights of neurons. We need to do this to avoid overfitting.
Read now
Unlock full access