Because their neurons receive all the values from the previous layer without any distinction (they are fully connected), these neural networks do not have a notion of distance/spatiality. Spatial relations in the data are lost. Multidimensional data, such as images, could also be anything from column vectors to dense layers because their operations do not take into account the data dimensionality nor the positions of input values. More precisely, this means that the notion of proximity between pixels is lost to fully connected (FC) layers, as all pixel values are combined by the layers with no regard for their original positions.
As it does not change the behavior of dense layers, to simplify their computations and parameter representations, it is common practice to flatten multidimensional inputs before passing them to these layers (that is, to reshape them into column vectors).
Intuitively, neural layers would be much smarter if they could take into account...
