Neural Network Explorer

This scene shows a genuine, correctly structured feedforward neural network — three layers, fully connected between adjacent layers, with a 3-node input layer, a 4-node hidden layer using the real ReLU activation function, and a 2-node output layer using the real softmax activation function. The weighted links you see, and the activation values shown when a layer is inspected, are computed live by actually running these standard activation functions on the network’s weights, the same real ReLU and softmax mathematics used throughout modern deep learning — not a decorative animation. The fixed example weights and input values are illustrative constants in a typical small random-initialisation range, not values from training a real model; the mathematics applied to them is what is real here.

Selecting a layer inspects it, highlighting its nodes and the values flowing into them. Edges are coloured by the sign of their weight — one colour for positive weights, another for negative — so you can see at a glance which connections excite or suppress the next layer’s activations. Unlike the other four scenes, this one has no guided camera tour: a standard left-to-right layered network diagram is legible from a single fixed framing, so free-look orbit navigation is offered instead.

Checking your browser’s 3D support…