论文笔记:Inception-V4, Inception-ResNet

1. Motivation: residual connections work well for deep network ==> can be combined with  Inception ( Inception-ResNet) 

 

2. architecture

a. replace filter concatenation of  inception with residual connection 

(the 1 x 1 conv after  inception layer aims to scale up the dimension before adding to the input)

b. scaling down the residuals (multiple scaling factor 0.1~0.3) before addition ==> stabilize the training (prevent weights from going to 0)

2. Batch normalization: on top of traditional layers (excluding summation layer to reduce computational cost), prevent saturating. 

3. performance.(for detailed differences bw models, please refer to the paper)