Sparsified Training of Convolutional Neural Networks

A device comprising:

• processing circuitry;
• a first memory configured to store image data, the image data including a plurality of blocks of image data for processing;
• and
• a second memory configured to store software executable by the processing circuitry for implementing a convolutional neural network (CNN) with a block multiply accumulation based (BMA-based) convolution kernel on the blocks of image data, in which the CNN includes:
• a plurality of convolution layers;
• and
• a plurality of filter coefficients including a respective subset of filter coefficients for each convolution layer of the plurality of convolution layers, each convolution layer associated with a threshold value;
• wherein a number of filter coefficients in a first subset of filter coefficients of a first convolution layer is different than a number of filter coefficients of a second subset of filter coefficients of a second convolution layer, and
• wherein a percentage of the coefficients of a third subset of filters in a third convolution layer has a value below a first threshold value, and a percentage of the coefficients of a fourth subset of filters in a fourth convolution layer has a value below a second threshold value, in which the first threshold value is different than the second threshold value.

A non-transitory device-readable medium storing software instructions that, when executed by one or more processors, implement a multi-layer convolutional neural network (CNN) with a block multiply accumulation based (BMA-based) convolution kernel on blocks of an input feature map, wherein, to implement the multi-layer CNN, the processor is configured to:

• for each layer of the multiple layers of the CNN,
• scan the layer to identify which filter coefficients of the layer are above a threshold value, the filter coefficients above the threshold value representing a group of executable filter coefficients;
• and
• for each executable filter coefficient of the layer, multiply each data element of a block of the input feature map corresponding to the executable filter coefficient of the layer by the executable filter coefficient and accumulate results into corresponding locations of a block of an output feature map, including skipping the multiply and accumulate operations for each filter coefficient of the layer below the threshold value;
• wherein the output feature map of a previously-processed layer becomes the input feature map for the next layer to be processed.