Question about paper

[King4819]

Hi, Can I ask in this plot of paper, the grid-like denotes kernel or the feature map? I think it should denotes feature map!

[janthmueller]

I believe that on $f_4^-$, the in_channels of a convolutional layer are pruned, which corresponds to the depth of all filters that the convolutional layer holds. On the other hand, $f_4^+$ provides a method to filter the output channels, equivalent to the number of filters used in the convolutional layer. As a reminder, the weight tensor of a convolutional layer in PyTorch has the shape (out_channels, in_channels, kernel_height, kernel_width). The paper explicitly states that $(:, k, :, :) \neq (k, :, :, :)$, indicating a difference between the pruning schemes. I believe they chose a 2D representation for simplicity, as it is more straightforward than representing a 4D tensor. Feature maps are not pruned directly; they are the result of pruning an entire filter, which aligns with the case on $f_4^+$.

[King4819]

@janthmueller Thanks for your reply, so you mean that these two figure corresponds to the same filters right, which are the filters in conv f4