the.com/dropout
the algorithm's version of not letting any single neuron become the office know-it-all.
means a regularization technique that randomly switches off neurons during training so a neural network can't over-rely on any one path.
from introduced by geoffrey hinton and colleagues around 2012, inspired by the idea that sexual reproduction shuffles genes precisely to stop overly specialized co-adapted traits from dominating.
typical rate20 to 50 percent of neurons silenced per pass
training onlyfully active again at test time
effectacts like training many thinner networks at once