[Paper note] Xception: Deep Learning with Depthwise Separable Convolutions

• paper

Intuition

• Inception series
• Conv maps cross-channel correlation and spatial correlation at the same time.
• Inception module makes this process easier and more efficient by explicitly factoring it into a series of operations that would independently look at cross-channel correlations and at spatial correlations.
• 1x1 Conv -> cross-channel correlation; 3x3 & 5x5 Conv -> spatial correlation.
  
• An extreme version of this separation is to entirely decouple the cross-channel and spatial operations, naming Xception.

Xception

• Xception module:
  
• First use 1x1 Conv
• Conduct depthwise separable convolution (DSC): each feature-map have different 3x3 Conv, then concatenate the result of each Conv.
• Advantages: Efficient parameter usage
• Whole model
  

Experiment

• Dataset: JET (internal Google dataset), ImageNet, FastEval14k.
• Result
  
  • Xception converges faster than Inception V3 and gets higher accuracy.
  • 21.0% top-1, 5.5% top-5 error on ImageNet.
  • Better with residual connection.
  • Worse with non-linear in between the 1x1 and DSC.