DSAverse

Initializing Sorting Algorithms...

Sorting

Trees

Graphs

Preparing interactive visualizations...

Back to Deep Learning Architectures

Convolution

A sliding dot product: the same nine weights laid over every patch of the picture in turn. Watch each multiply-add, then edit the kernel, stride and padding yourself.

Builds onDot Product Forward Pass

Hand-made 7×7 pictures and classic kernels · every sum computed live

Example:

Act 1 of 7 · Picturestep 1 of 15

picture [H, W] =[7, 7]

kernel [3, 3]

feature map [H′, W′] =[5, 5]

Speed

To a network a picture is a grid of numbers: 7 × 7 brightness values from 0 (dark) to 3 (bright). Nothing about it says "edge" or "line" — a convolution is how the network finds those.

y(i, j) = \sum_{a=0}^{2}\sum_{b=0}^{2} w(a, b)\; x(i + a,\, j + b)

Convolution

A sliding dot product: the same nine weights laid over every patch of the picture in turn. Watch each multiply-add, then edit the kernel, stride and padding yourself.

Builds onDot Product Forward Pass

Hand-made 7×7 pictures and classic kernels · every sum computed live

Example:

Act 1 of 7 · Picturestep 1 of 15

picture [H, W] =[7, 7]

kernel [3, 3]

feature map [H′, W′] =[5, 5]

Speed

To a network a picture is a grid of numbers: 7 × 7 brightness values from 0 (dark) to 3 (bright). Nothing about it says "edge" or "line" — a convolution is how the network finds those.

y(i, j) = \sum_{a=0}^{2}\sum_{b=0}^{2} w(a, b)\; x(i + a,\, j + b)

DSAverse

Convolution

See also

Convolution

See also