At test time dropout is approximated by
AInverting the mask used at the start of training only
BSampling many random dropout masks and averaging
CMultiplying each output by 0.5 with no neuron zeroed
DDisabling all hidden neurons and using only FC layers
Answer & Solution
Correct answer: C. Multiplying each output by 0.5 with no neuron zeroed
At test time all neurons are used, but their outputs are multiplied by 0.5 (the keep probability). This approximates the geometric mean of the predictive distributions over the exponentially many dropout sub-networks.
Related questions
Image similarity in AlexNet's last hidden layer is measured byThe SGD update rule for AlexNet usesRemoving any single middle convolutional layer of AlexNet costs roughlyAlexNet was trained for roughly how many epochs on ImageNet?AlexNet uses overlapping pooling withAlexNet's second form of data augmentation isAlexNet's first form of data augmentation isWhy are ReLU networks faster to train than tanh networks?