What do the two Subtract operations mean?

[CacatuaAlan]

Hi! Mentioned that your paper has two Subtract operations which confuses me. Can I just consider them as decouple?

[Lupin1998]

Hi, @CacatuaAlan, sorry for the late reply! You can regard the two subtract operations in the feature decomposition (FD) and the channel aggregation (CA) modules as removing the low-pass components. Then, we adaptively combine the low-pass and the rest components to increase the diversity (i.e., enhancing the modeling of middle-order interactions). Feel free to ask me when new problems occur and star our repo if it's helpful to your project!

[CacatuaAlan]

Thank you for your patient explanation. In the two subtraction operations within the FD and CA modules, the subtracted components are different operations (GAP and conv1x1). Is there any special design or experimental comparison for this?

Hi, @CacatuaAlan, sorry for the late reply! You can regard the two subtract operations in the feature decomposition (FD) and the channel aggregation (CA) modules as removing the low-pass components. Then, we adaptively combine the low-pass and the rest components to increase the diversity (i.e., enhancing the modeling of middle-order interactions). Feel free to ask me when new problems occur and star our repo if it's helpful to your project!

Thank you for your patient explanation. In the two subtraction operations within the FD and CA modules, the subtracted components are different operations (GAP and conv1x1). Is there any special design or experimental comparison for this?

[Lupin1998]

Well, from my perspective, I try to regard the two subtraction operations in the FD and CA modules with a similar intuition, i.e., removing the low-order features (so-called DC component or low-frequency component) and reweighting them. As for ablation experiments, I have found the two subtractions useful. If you are interested in whether we could achieve this intuition with a different operation, I think the addition might be supported by a similar operation, GRN (proposed in ConvNeXt.V2).