Dropout is a widely used technique to improve convergence and robustness of a neural net and prevent neural nets from overfitting. It consists of setting some random values to zero for the layers on which we'd like it to apply. It introduces some randomness in the data at every epoch.
Usually, dropout is used before the fully connected layers and not used very often in convolutional layers. Let's add the following lines before each of our two fully connected layers:
dropout = 0.5 if dropout > 0 : mask = srng.binomial(n=1, p=1-dropout, size=hidden_input.shape) # The cast is important because # int * float32 = float64 which make execution slower hidden_input = hidden_input * T.cast(mask, theano.config.floatX)
The full script is in 5-cnn-with-dropout.py. After 1,000 iterations, the validation error of the CNN with dropout continues to drops down to 1.08%, while the validation error of the CNN without dropout will not go down by 1.22%.
Readers who would like to go further...