Writing

ExplorableCNNs

A filter slides over an image

Move a 5×5 filter across a 32×32×3 image and see exactly which pixels feed each cell of the activation map.

 

Reading the scene: more ink on a sheet means a larger value, and the map is shaded relative to its own largest value. On the filter face, blue is a positive weight and orange a negative one. Simplified: each preset repeats one F×F pattern on all three channels at one third weight, the bias is 0, and padding adds zeros.

What to look for

  • One filter, one map. The same weights are reused at every stop, so the parameter count depends on F and the number of filters and never on the image size. This page shows one filter. A layer with six filters stacks six such maps.
  • ReLU keeps one polarity. With the vertical edge filter, the left side of each shape lights up. The right side gives a negative sum, and ReLU clips it to zero.
  • Shift in, shift out. The map keeps the layout of the image: the square sits top left in both. Convolution is translation equivariant. The rough invariance to small shifts comes later, from pooling.