Spatial correlation is the tendency for nearby positions in structured data to be related. In images, neighboring pixels often share colors or shades, while distant pixels may have little direct relationship. Shuffling pixels destroys the image’s meaning because position carries information. section 3.1 section 3.2
Convolutional neural networks encode this prior by treating nearby items as related and far-apart items as largely unrelated. A convolution combines an input image with a filter to highlight patterns, such as a horizontal stroke. section 3.2
A digit image
In an image of the digit 5, a black pixel usually has neighboring black pixels, and a white pixel usually has white or nearly white neighbors. A filter can detect a local feature such as the top horizontal line of a 7. section 3.2
