Enter input values and watch them flow through a small neural network — every weighted sum and activation shown live.
A forward pass is how a neural network turns input data into a prediction, data flows through the network one layer at a time, left to right, with no looping back. At each neuron, the process is identical: multiply each incoming value by its corresponding weight, sum all those products together plus a bias term, then pass that sum through an activation function to produce the neuron's output. This same simple computation, repeated across every neuron in every layer, is the entire mechanism behind how a trained neural network makes predictions, whether it's a tiny 3-neuron network like this one or a modern LLM with billions of parameters.
The edges in the diagram above are colored green for positive weights and red for negative weights, with thickness reflecting magnitude. These weights are exactly what "training" a neural network means, adjusting these numbers (via backpropagation and gradient descent) so the network's outputs become more accurate over time. A network with random, untrained weights (try the "Randomize Weights" button) produces essentially meaningless output, training is the process of finding weight values that make the output actually useful for the task at hand.
Without an activation function, a neural network, no matter how many layers deep, mathematically collapses into being equivalent to a single linear transformation, stacking linear operations without nonlinearity in between just produces another linear operation. Activation functions introduce nonlinearity at each layer, which is precisely what lets neural networks learn complex, curved decision boundaries and represent relationships far beyond what a straight line or flat plane could capture. This is why "activation function" isn't an optional add-on, it's a structural requirement for a neural network to be more expressive than plain linear regression.
The trace below the diagram shows the exact arithmetic happening at each neuron, the weighted sum calculation followed by the activation function applied to it. Watching these real numbers update as you change the inputs is a genuinely effective way to build intuition for what's actually happening inside a network, rather than treating it as an unexplainable black box, every single number flowing through this tiny network is fully visible and traceable.