🧠 Neural Network Architectures

About this simulation

A neural network is built from layers of simple units ("neurons") connected by weighted links. What differs between architectures is how those units are wired together — and that wiring pattern is exactly what this scene lets you swap between and watch in 3D.

How it works

• Feedforward (MLP) — every neuron in one layer connects densely to every neuron in the next; signals flow strictly forward, layer by layer.
• Convolutional (CNN) — each neuron only connects to a small local window in the previous layer (its receptive field), and that same small filter slides across the whole layer — shown here as a glowing kernel sweeping down the stack.
• Recurrent (RNN) — columns represent the same hidden units unrolled across time steps; a strong horizontal chain carries hidden state forward, and small loops show each unit feeding its own next state.
• Transformer (attention) — every token attends to every other token directly, regardless of distance, producing a dense, all-to-all web of connections whose strength pulses over time.

Controls

• Architecture — switch the connectivity pattern between the four families above.
• Layers / Neurons per layer — resize the network (rebuilds automatically).
• Signal speed — how fast the glowing pulses travel along connections during the forward pass.
• Show weight strength — toggle whether connection brightness reflects a randomly assigned weight.

Did you know?

Convolutional filters reuse the same small set of weights at every position in an image, which is why CNNs need far fewer parameters than a fully-connected network of the same size — and why transformers, which drop locality entirely, needed much larger datasets before they became practical.