Question 773

Same Stencil, Every Pixel

A convolutional layer slides the same 3x3 filter across an entire input image to produce a feature map. Compared to a fully connected layer with one weight per input pixel, what two properties does this convolution operation exploit that make it well suited to image data?

A

Random weight initialization at each spatial position and full connectivity between all pixels

B

Batch normalization across channels and dropout between layers

C

Weight sharing (the same filter weights are reused at every spatial position) and locality (each output depends only on a small neighborhood of input pixels)

D

Global attention (every output pixel attends to all input pixels) and dynamic weight generation per input

E

Recurrent weight reuse across time steps and gated memory cells