Neural Network Forward Pass: Live Activation Math (2D)

The 3D version shows five layers of neurons exchanging decorative light pulses on a timer — visually pleasing, but the pulses aren't tied to any real computation. This 2D companion replaces that with an actual 3–5–5–3 feedforward network: every wave is a genuine forward pass, computed from Xavier-initialized weights, an adjustable bias offset and whichever activation function you pick (sigmoid, ReLU or tanh). Drag the three input sliders and the softmax output bars update from real matrix arithmetic, not a script; the pulse colour along each synapse even encodes the sign of that connection's actual weighted contribution, so you can watch positive and negative signals fight for the winning class in real time.