Turn the dials, watch a single neuron think. Signals flow, the sum builds, and the neuron decides to fire — or stay quiet.
Inputs & Weights
1.00
1.00
0.50
0.50
-0.70
The Neuron
· quiet ·
Edge thickness = weight sizecyan = positive · coral = negativePulses speed up with stronger signal
Firing Rule
Live Math
What it does for every 0/1 combo — the neuron's "truth table", updating live
CHALLENGE 1 OF 4
Build an AND gate
Set the dials so the live truth table matches the target on the right, then hit Check.
Target — fire (⚡) exactly here
The panels above are your live neuron. Match it to this.
That's the XOR wall — and it's not your fault.
One neuron can only draw one straight line to split "fire" from "quiet". AND and OR are splittable by a line. XOR is not — the two "fire" cases sit on opposite corners. No weights can fix it.
The fix from the slides: add a hidden layer. Two neurons each draw a line; a third combines them. Toggle the inputs below and watch a 2‑2‑1 network nail XOR.
Backprop Trainer — teach a 2-layer net to solve XOR
4
0.80
Hidden activation
forward pass backward pass
Press Train and watch the loss fall as backprop nudges every weight toward the answer.
epoch 0loss —
Decision boundary (prediction across the whole input square)
Each bar = how strong the gradient still is when it reaches that layer during backprop (output = tallest). Watch it fade toward the input — or blow up.
Healthy
From the slides: a vanishing gradient is a whisper down a long line — too faint by the time it reaches the front. Exploding is a tap on full blast, overflowing the cup. Fixes: ReLU, careful initialisation, gradient clipping, batch-norm, and skip-connections.
Built as a companion to “A Soft Introduction to Artificial Neural Networks”. Tweak any dial — nothing here is fixed in stone. no install needed