stem/AI/Neural Networks/CNN/FCN/ResNet.md

- Residual networks
- 152 layers
- Skips every two layers
	- Residual block
- Later layers learning the identity function
	- Skips help
	- Deep network should be at least as good as shallower one by allowing some layers to do very little
- Vanishing gradient
	- Allows shortcut paths for gradients
- Accuracy saturation
	- Adding more layers to suitably deep network increases training error

# Design

- Skips across pairs of [[Convolutional Layer|conv layers]]
	- Elementwise addition
- All layer 3x3 kernel
- Spatial size halves each layer
- Filters doubles each layer
- [[FCN|Fully convolutional]]
	- No fc layer
	- No [[Max Pooling|pooling]]
		- Except at end
	- No dropout

![[imagenet-error.png]]

![[resnet-arch.png]]
![[resnet-arch2.png]]