สไตรด์ แพดดิง และพูลลิง

When a 3x3 kernel slides over an image, the edges have no room. So the output feature map comes out a little smaller than the input.

Padding wraps the image in a ring of zeros so the kernel can reach the corners. Now the output can stay the same size as the input.

For a 3x3 kernel, padding of 1 keeps the spatial size unchanged. People call this same padding.

Stride is how far the kernel jumps each move. A stride of 1 visits every position; a larger stride skips ahead.

With stride 2 the kernel hops two pixels at a time, so the output is roughly half the width and height. It is a cheap way to downsample. ⬇️

Combine them to control the exact output size. Here stride 2 with padding 1 downsamples while keeping the corners covered.

Shrinking the maps cuts computation and lets later layers see a wider region of the original image with each kernel.

Pooling downsamples too, but with no learned weights. It just summarizes each small patch into a single value.

Max pooling takes the largest value in each 2x2 patch. It keeps the strongest signal and quietly drops the rest.

Average pooling takes the mean of each patch instead of the max, giving a smoother, gentler summary of the region.

Because pooling summarizes a patch, a tiny shift in the input barely changes the output. This gives a touch of translation invariance.

Let us test what padding actually does for a convolution.

You learned that padding preserves size, stride sets the step and downsamples, and pooling shrinks maps while adding a little shift tolerance. 🎯