vggnet

VGGNet, from the Visual Geometry Group at Oxford (Simonyan and Zisserman, 2014), is the architecture that answered a clean question: if you keep everything as uniform and simple as possible and just make the network deeper, how far does accuracy improve? Its answer was a network built almost entirely from one repeated motif, the 3x3 convolution with stride 1 and padding 1, interleaved with 2x2 max pooling to halve resolution, stacked 16 or 19 weight layers deep (giving the famous VGG-16 and VGG-19). There are no exotic modules; the appeal is precisely its monotonous regularity, which made it the default feature extractor and teaching example for years.

The central technical argument of VGG is that a stack of small 3x3 convolutions is strictly preferable to a single large convolution with the same receptive field. Two stacked 3x3 layers see the same 5x5 input region as one 5x5 layer, and three stacked 3x3 layers cover a 7x7 region, but the stack uses fewer parameters (for the same channel count, three 3x3 layers cost 27 channel-pairs of weights versus 49 for one 7x7) and inserts two extra nonlinearities, making the learned function more expressive. This insight, that depth of small filters beats width of large filters, became a permanent design principle.

VGG's weakness is its cost. Because it keeps three large fully-connected layers at the end (two of 4096 units feeding a 1000-way classifier), VGG-16 carries roughly 138 million parameters, the vast majority of them in those final dense layers, and it is computationally heavy to run. Later networks such as GoogLeNet and ResNet matched or beat its accuracy with a small fraction of the parameters by replacing the fully-connected stack with global average pooling and by using residual connections. Even so, VGG remains widely used as a fixed pretrained backbone, especially as a perceptual feature space for tasks like style transfer and image-similarity losses, because its features are clean and well understood.