Input · 96 × 96draw with your pointer
Kernel · 3 × 3drag cells ↕ · double-click = 0
each output pixel = Σ (kernel × 3×3 patch) — exactly what a CNN computes, thousands of times per layer. Edit a cell and this becomes your filter. In a trained network, backprop finds these numbers on its own.
Pipeline
Filter bank · one conv layer, six filterswhat different kernels notice