A U-Net turns an image into another image of the same size. This one is as small as a U-Net can be: 8×8 pictures, 207 weights. Small enough to see every number.
The shape. The encoder (left) looks at the picture through 3×3 windows. Max-pooling halves the resolution, so the bottleneck (bottom) sees the whole picture coarsely. As in real U-Nets, the channels double where the resolution halves: 2 at full resolution, 4 in the bottleneck. A learned 2×2 up-convolution brings it back up to 2 channels at full resolution, and the skip connection hands the decoder (right) the encoder’s full-resolution map as well: coarse context from the bottom, fine detail from across.
Step through it. ➡️ step does one stage: the forward pass layer by layer, then the backward pass layer by layer, then the weight update. Tap any neuron or weight to see the exact arithmetic behind it. With 📍 follow on, each step selects a neuron it just computed.
Move a weight yourself. Tap a weight or bias, then drag the ✏️ slider. The current sample’s forward pass reruns with the new value, so every map, the output, the loss and the test images respond as you drag. ↩️ puts it back. On a wide screen, drag the bars between the panels to resize them.
See why it is shaped like a U. Pick a task, ▶️ play until it learns, then switch off one of the two paths into the decoder, press 🔄 new weights, and train again. 🔲 outline needs the skip connection; 🔍 big only needs the bottleneck.
If it stalls at 89% on 🔲 outline (about one start in three), the network has learned to copy its input. That is most of the way to an outline, and almost no gradient points away from it. Press 🔄 new weights for a fresh start.