3D Neural Network: Signal Flow

Learning rate η 0.50

Hidden neurons 6

Activation function

Epoch

Loss

–

Accuracy

–

Drag to orbit, scroll to zoom. Each sphere is a neuron — its glow is its activation. Glowing pulses travel input→hidden→output along the actual weighted connections every step, showing the literal forward pass through space; edge color/thickness still encodes each weight's sign and size.

How it works

Same network as the formula from the article, laid out as physical layers instead of a flat diagram:

aₕ = f(W₁·x + b₁)
ŷ  = σ(W₂·aₕ + b₂)

Every training step is one pass of gradient descent on the cross-entropy loss:

Δw = η · ∂Loss/∂w

found via backpropagation. Here the forward pass isn't just computed — it's animated as particles moving along the 3D edges, layer by layer, so the "signal flowing through the network" metaphor is literal geometry instead of an abstract diagram.

Weights: edge color/thickness, updated every step by backprop.
Bias: baked into each neuron's activation, not drawn separately.
Rotate the camera to see the input layer (2 nodes), hidden layer (ring) and single output neuron stacked along depth.

positive weight / high activation negative weight / low activation

⚠️ WebGL failed to load.

Try a different browser or enable hardware acceleration.