Lupin1998
commented
1 year ago Hi, @123456789asdfjkl, thanks for your excellent question! Unfortunately, there is also a typo, and it should be 0-order interaction of each patch itself and 'n-order' interaction covering all patches. Given an image of n patches, the m-order interactions are defined by Eq.1 and Eq. 2 in Representation Bottleneck, where $0\ge m\le n-2$. There are two trivial conditions: (a) When the interaction pair i,j is the same patch, it will be 0-order interaction that the patch interacts with itself, i.e., $Conv{1\times 1}(\cdot)$. (b) When all patches are pooled into a single token, e.g., by global average pooling GAP($\cdot$), it cannot be measured by the concept of m-order interaction (because the contextual set cannot be larger than m-2). Therefore, modelling two trivial conditions by $Conv{1\times 1}(\cdot)$ and GAP($\cdot$), the proposed FD module can intend to help the model learn more useful interactions.
123456789asdfjkl
Hi！Thank you for your great work！我对高亮的这段话没能理解，想请教您，这个应该如何理解