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

29 lines
747 B
Markdown
Raw Normal View History

- 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]]