Receptive Field

• Not for Dense, only local connected Layers like Conv, Pooling
• A neuron’s receptive field is the patch of the total field of view that a single neuron has access to
• Almost a logarithmic relationship between classification accuracy and receptive field size
  • Large fields are almost necessary for high level recognition tasks, but with diminishing rewards
• $r(i-1)=s_i\times r_i+(k_i-s_i)$
• 
• Recursive
  • $r_0={\Sigma }_{i=1}^L((k_i-1){\Pi }_{j=1}^{l-1}{s}_j)+1$
• How to increase receptive field
  • Add more Conv
  • Add Pooling or higher stride
  • Causal Dilated Conv
  • Depthwise Separable