Efficient convolutional neural networks and techniques to reduce associated computational costs

A computing system comprising:

• one or more processors;
• and
• at least one tangible, non-transitory computer-readable medium that stores a convolutional neural network implemented by the one or more processors, the convolutional neural network comprising:
• a plurality of convolutional blocks, each of the plurality of convolutional blocks configured to receive an input and generate an output, at least one of the plurality of convolutional blocks comprising:
• a projection separable convolutional layer configured to apply a depthwise convolution and a pointwise convolution during processing of an input to the projection separable convolutional layer to generate an output of the projection separable convolutional layer, and wherein the output of the projection separable convolutional layer has a depth dimension that is less than a depth dimension of the input of the projection separable convolutional layer;
• an activation layer configured to receive the output of the projection separable convolutional layer and generate an input for an expansion separable convolutional layer;
• the expansion separable convolutional layer configured to apply a depthwise convolution and a pointwise convolution during processing of the input for the expansion separable convolutional layer to generate an output of the expansion separable convolutional layer, and wherein the output of the expansion separable convolutional layer has a depth dimension that is greater than a depth dimension of the input of the expansion separable convolutional layer, and wherein the depthwise convolution of at least one of the projection separable convolutional layer or the expansion separable convolutional layer is applied with a kernel size that is greater than 3×3;
• and
• a residual shortcut connection from the input of the projection separable convolutional layer to the output of expansion separable convolutional layer.