Convolutional Networks · L2 Mechanics

The mechanics of convolution, stride, padding, pooling, and how spatial dimensions shrink through a network.

Input

32×32×3

After conv (3×3, stride 1)

30×30×16

After pool (2×2)

15×15×16

After conv (3×3, stride 1)

13×13×32

Spatial dimensions shrink while channel depth grows as the network gets deeper.

A convolution with a $k \times k$ filter slides across the input, computing a dot product between the filter and each local patch it covers. Three parameters control the geometry: kernel size (filter dimensions), stride (how many pixels to move each step), and padding (zeros added around the border).