Node Specificity in Convolutional Deep Nets Depends on Receptive Field Position and Size
This addresses a fundamental issue in convolutional neural networks for computer vision researchers, revealing limitations in positional invariance that could impact model design and interpretation.
The paper investigates how receptive field (RF) size and position in convolutional deep neural networks affect node specificity, showing that as RFs approach full image coverage, different positions lead to varying specificity due to portions falling outside the input space, which challenges positional invariance and enables complex context specificity.
In convolutional deep neural networks, receptive field (RF) size increases with hierarchical depth. When RF size approaches full coverage of the input image, different RF positions result in RFs with different specificity, as portions of the RF fall out of the input space. This leads to a departure from the convolutional concept of positional invariance and opens the possibility for complex forms of context specificity.