> In the past few years Convolutional Neural Networks (CNNs) have revolutionized the field of computer vision. Each year the ImageNet Challenge (ILSVRC) has seen plummeting error rates due to the ubiquitous adoption of CNN models amongst the contestants.
Am I right that the "Convolution" part only refers to the speed by which the models can be trained, and not to any other quality of these models?
No. Most of layers in a CNN perform convolutions with kernels. This is not the same as standard DNNs that do a full matrix multiply.
Convolutional kernels allow you to use many fewer variables to perform the forward layer operation; and CNNs tie these trainable variables across layers. Training is not only faster, but also more robust because you have less parameters to learn.
Am I right that the "Convolution" part only refers to the speed by which the models can be trained, and not to any other quality of these models?