Why use the pool(x) - x

[FlyEgle]

I have a quesition for use the pool(x) - x

I give the answer is : pooling is used for smooth x, '-' mean model ignore the same value with the pool size and focus the neightbor wihch small or bigger than smooth x. Reinforcement the relation for the 8-neighboor.

So this pooling is the local-relation, than use the conv mlp mixer to exchange the information for gloabl.

Is that reason? Thank for you replay

[yuweihao]

Hi @FlyEgle ,

As shown in the paper, since the MetaFormer block already has a residual connection, subtraction of the input
itself is added in Equation (4).

[wsy-yjys]

@yuweihao Hi, have you tried pool(x)+x or pool(x)?

[yuweihao]

Hi @wsy-yjys , I tried pool(x) and pool(x) - x. I remember pool(x) -x is better than poo(x) by about 0.2 or 0.3 accuracy (the number is a bit blurry). I have not tried pool(x) + x.

[wsy-yjys]

@yuweihao Thank you for your reply. I am very curious how did you come up with the idea of pool(x) - x at that time?

[yuweihao]

@wsy-yjys , from the perspective of image processing, pool(x) is Mean Filter while pool(x) -x ` is Laplacian Filter. At that time, I wanted to specify the token mixers as extremely sample operators, so I tried these simple filters without any learnable parameters. I also gave another explanation in this issue #4 (comment)