Local Connectivity
A fully connected layer can connect every output to every input.
A convolutional layer usually connects each output to a local region. For an image, that region is a small patch. For a sequence, it is a short window.
Local connectivity matches a useful assumption: nearby values often matter together.
For example, an edge in an image is local. It depends on neighboring pixels, not every pixel in the image at once.
This does not mean global context is unimportant. It means early layers can build local features first. Deeper layers can combine those features into wider patterns.
A convolution reads a 3 x 3 image patch. How many input positions are in the patch?
Compute it first, then check your number.
Enter 1 if local connectivity reads a small neighborhood, or 2 if it always reads the entire input grid.
Compute it first, then check your number.