🧠 Inside a Neural Network: Forward Pass

About this simulation

A neural network turns raw numbers into a prediction through a forward pass: each layer takes the activations of the layer before it, combines them with learned weights and a bias into a weighted sum, and squashes that sum through an activation function to produce the next layer's activations. Repeating this from the input layer to the output layer produces the network's final prediction.

How it works

• Each sphere is a neuron; its glow brightness encodes its current activation value (0 = dim, 1 = bright).
• Each line is a weighted connection; green lines carry positive weights, red lines carry negative weights, and line thickness/opacity scales with |weight|.
• Running the pass sends a pulse along every connection into the next layer, then that layer's neurons compute z = Σ(weight × input) + bias and apply the chosen activation function before lighting up.
• The two output neurons are the network's class scores; the brighter one is the predicted class.

Controls

• Input 1–3 — the raw feature values fed into the input layer.
• Activation — the nonlinearity applied at every hidden and output neuron (ReLU, Sigmoid, or Tanh).
• Weight scale — multiplies the magnitude of the network's random weights, showing how larger weights make activations swing more sharply.
• Pass speed — how fast the pulse animation moves from layer to layer.
• ▶ Run forward pass — replays the animation with the current inputs and weights.
• ↺ New random network — draws a fresh set of random weights and biases.

Did you know?

Without a nonlinear activation function, stacking any number of layers would collapse mathematically into a single linear transformation — activation functions like ReLU are what let deep networks approximate curved, complex decision boundaries at all.