M17 / LABORATORY

Training MLP. Watch the network
learn, neuron by neuron.

The MLP module showed one training step in numbers. Here you watch thousands of them as a living diagram: edges thicken and flip colour as weights update, neurons light up as signal flows forward, the error pulses backward, and a decision boundary forms in real time. Same backprop — now visual.

EXPERIMENT 17.01 · MLP.LIVE EPOCH · 0 LOSS · — ACC · —

DATASET

ARCHITECTURE

HIDDEN H₁ 6

HIDDEN H₂ 4

ACTIVATION

OPTIMIZATION

LR η 0.10

SPEED 6

Hover the boundary plot → the network shows that point's forward pass.

NETWORK · edge colour = weight sign, thickness = |weight| idle

▬ +weight ▬ −weight ● activation backward pulse = gradient

DECISION BOUNDARY

LOSS

What to watch → At init, edges are thin and random-coloured, the boundary is noise. As training runs, certain edges thicken (weights that matter grow), some flip sign, hidden neurons specialize, and the boundary snaps to the data. Try SPIRAL with too few hidden units — it physically can't bend enough, and the loss plateaus. Add units and it suddenly fits. That's capacity, live.