From time to time, it’s nice to revisit history. VGGNet is only 4 years old (submitted to ImageNet Challenge 2014) yet it was surpassed many times by ResNet. Even though ResNet is both more accurate and almost 4 times faster, VGGNet sees new usage thanks to it’s 2 fully connected feature layers.
Those dense layers found application in various tasks as they generalize well to other problems ‒ image embedding, taking part in other networks or even replacement for L2 loss as a perceptual loss (full paper).
Full paper: Very Deep Convolutional Networks for Large-Scale Image Recognition
Keras implementation: github.com/simongeek
Tensorflow implementation: github/machrisaa
