VGG

VGG demonstrated that uniformity and depth — rather than bespoke layer designs — were the dominant factors driving image- classification accuracy in 2014. The network is built from a single recipe: stack $3\times3$ convolutions with stride 1 and padding 1, periodically halve the spatial resolution with a $2\times2$ max-pool, double the channel count at each pooling boundary, and end with three large fully-connected layers.

The $3\times3$ kernel choice is theoretically motivated. Two stacked $3\times3$ convolutions have the same receptive field as a single $5\times5$ convolution but use only $2 \cdot 9 C^2 = 18 C^2$ parameters versus $25 C^2$, while inserting an extra nonlinearity between them. Three stacked $3\times3$ layers match a $7\times7$ receptive field at $27 C^2$ versus $49 C^2$ parameters and two extra nonlinearities. Depth at a fixed receptive field therefore buys representational power essentially for free.

The paper's variants (configurations A, B, D, E corresponding to VGG-11/13/16/19) differ only in how many $3\times3$ convs are stacked between pooling layers. Despite a parameter count of 138 M — most of it locked up in the 4096-dim fully-connected layers — VGG-16 achieves a top-5 ImageNet error of 7.3%, second only to GoogLeNet in ILSVRC-2014, and its features transferred so well that "VGG features" became the standard backbone for downstream tasks (object detection, segmentation, style transfer) for years after publication.