VGGNet
A deep CNN that demonstrated the effectiveness of using very small (3x3) convolution filters.
Overview
VGGNet, developed by the Visual Geometry Group at Oxford in 2014, demonstrated that network depth with small convolution filters (3x3) is critical for good performance in image recognition. VGG-16 and VGG-19 (with 16 and 19 weight layers respectively) achieved state-of-the-art results on ImageNet.
Key Details
VGGNet's uniform architecture — stacking 3x3 convolutions with max pooling — made it simple to understand and implement. While it has been surpassed by ResNet and other architectures, VGGNet remains widely used as a feature extractor and for style transfer applications. Its main drawback is the large number of parameters (138M for VGG-16).