CNN explorer, in 3D
The same toy network as the 2D page, drawn as what it is: a sequence of H×W×C blocks. Depth is channels. Cube size is magnitude, colour is sign (blue negative, orange positive). Drag to orbit, scroll to zoom, right-drag to pan, and use the bar under the view to travel along the network. "Side by side" lays the stages out left to right; "depth" sends them away from you along the data-flow axis and draws the connections between stages, so the funnel from 16×16 down to 10 numbers is visible as a shape. Hover any cube for its receptive field; use the sweep controls to watch a kernel move through a block — the layer-2 kernel is a 3×3×4 cube that eats all four channels at once.
Hover a cube.
startend
Input
pool
Channel spacing
0 = solid H×W×C block, as the maths sees it. Larger = exploded into separate feature maps, as the 2D page draws them.
Sweep a kernel
Layer 1: a 3×3×1 window slides over the input and writes one cube of one output channel. Layer 2: a 3×3×4 window slides over the whole pool-1 block, and its four channel partial sums add into one cube.
Weights
Layer-1 kernels are Sobel-x, Sobel-y, Laplacian, diagonal. Layer 2 and FC are random: the prediction is arbitrary, the mechanics are real.
kernel window (forward) receptive field (backward)
Terms
Everything on this page is built from these. Hover or tap a dotted term above for the same definition in place.