Acceleration of convolutional neural network training using stochastic perforation

A computer implemented method to train a convolutional neural network, the method comprising:

• receiving training data to train the convolutional neural network using one or more convolutional layers, in which each convolutional layer comprises input maps, convolutional kernels, and output maps;
• creating a first perforation map for optimizing a convolution evaluation by filtering calculations used to generate the output maps using a forward pass of a convolutional layer, the forward pass comprising a first set of convolution calculations of the input maps and the convolutional kernels, wherein for filtering the calculations the convolution calculations are selectively performed based on the first perforation map, the first perforation map indicates a first subset of the first set of convolution calculations to be performed and a second subset of the first set of convolution calculations to be skipped, the first perforation map is stochastically generated;
• create a second perforation map for optimizing the convolution evaluation by filtering calculations used to determine gradients using a backward pass of the convolutional layer, the gradients determined based on a training error and a cost function associated with the convolutional neural network, the backward pass comprising a second set of convolution calculations of the output maps and the convolutional kernels, and wherein for filtering the calculations the convolution calculations are selectively performed based on the second perforation map, the second perforation map indicates a first subset of the second set of convolution calculations to be performed and a second subset of the second set of convolution calculations to be skipped, the second perforation map is stochastically generated, and the second perforation map being distinct from the first perforation map;
• create a third perforation map for optimizing the convolution evaluation by filtering calculations used to update the convolutional kernels using an update pass of the convolutional layer, the update pass comprising a third set of convolution calculations of the input maps and output maps, wherein for filtering the calculations the convolution calculations are selectively performed based on the third perforation map, the third perforation map indicates a first subset of the third set of convolution calculations to be performed and a second subset of the third set of convolution calculations to be skipped, the third perforation map is stochastically generated, and the third perforation map being distinct from the first perforation map and the second perforation map;
• and
• iteratively repeating the forward pass, the backward pass, and the update pass until the convolutional neural network is trained, the first perforation map, the second perforation map, and the third perforation map being different for each iteration.