Skip to content

LOCAL RESPONSE NORMALIZATION FOR FULLY CONNECTED

Jump to bottom
Riccardo Viviano edited this page Apr 11, 2021 · 2 revisions

A fully connected layer can be seen as a convolutional layer where each kernel has the dimensions:

  • channels := 1
  • kernel_rows := input_dimension
  • kernel_cols := 1

The number of kernels can be seen as the output of the fully connected layer:

  • n_kernels := output_dimension

For this reason i decided to use the local response normalization also for fully connected layers. (Even if the effect of layer normalization has no been proved, this kind of normalization is the only normalization that can be allowed during the edge popup algorithm, because there is no normalization parameters that must be learned)

Clone this wiki locally