Sliding Pattern Detectors

Reuse one detector everywhere

A convolutional neural network (CNN) applies small learned grids called filters or kernels across an image. At each location, it multiplies overlapping pixel values by kernel weights and sums them, producing one value in a feature map. The same kernel is reused at every location, a property called weight sharing.

image patch          vertical-edge kernel
[1 0 0]              [ 1  0 -1]
[1 0 0]      dot     [ 1  0 -1]  -> strong positive response
[1 0 0]              [ 1  0 -1]

The stride is how many pixels the kernel moves each step. Padding adds border values so edge pixels receive comparable coverage and output size can be controlled. Early filters often respond to local edges or textures; later layers combine maps into larger patterns.

Analogy: Slide a small stencil across a photograph. At every position, record how strongly the local patch matches the stencil. One stencil can find the same edge anywhere in the image.

Note: This is a numeric thought experiment only. No PyTorch, TensorFlow, or Keras dependency is available in the offline lab image.