Backpropagation — Watch Gradients Flow Through an MLP (2D)

Network

Hidden width H 5

Activation

Output

Softmax (classification) Linear (regression)

Learning rate η 0.10

Dataset

Mode

Live Readout

Loss

—

Accuracy

—

Epoch

|∇|

—

Layers

2→5→5→2

Parameters

Click the plane (left) to add a class-0 point (red), right-click for class-1 (blue). Backward pass shows δ pulses flowing right→left; edges glow by |∂L/∂w|.

How it works

Each backward step applies the chain rule layer by layer: δ^(l) = (W^(l+1))^Tδ^(l+1) ⋅ σ'(z^(l)), then ∂L/∂W^(l) = δ^(l)(a^(l-1))^T. Weights update as W ← W − η∂L/∂W. The same math as the 3D version — this view renders it with plain Canvas2D instead of WebGL.