Index page

VGG net

[2] VGG net Simonyan, Karen, and Andrew Zisserman. "Very deep convolutional networks for large-scale image recognition." arXiv preprint arXiv:1409.1556 (2014). Paper

Notes

Part1

• max pool: 2, 2
• conv: stri = 1, pad =1 ,ks = 3
• Always add ReLU
• Not all layers have max pool
• Adding 1x1 layers increases non linearity

  Network in network - 1x1

• mom = .9, decay = 5*10^-4, dropout = 0.5
• learning rate decay was used
• Improve convergence
• Implicit regularization -> greater depth, smaller filter
• Pre init of some layers
• Adding small amounts of noise to image -> increases accuracy
• Fusion
• Averaging best soft max parts of multiple performing models

Part2

• Object localization
• Euclidean loss -> penalizes deviation of bounding box
• No scale jittering