✂️ Neural Network Pruning Explained

About this simulation

A trained neural network typically has far more weights than it strictly needs to reach its accuracy — many connections end up with a magnitude near zero and contribute almost nothing to the output. Pruning identifies and removes these low-value weights (or entire neurons), shrinking the model and speeding up inference with little to no loss in accuracy, up to a point.

How it works

• Unstructured pruning removes individual weights below a magnitude threshold, leaving an irregular, sparse pattern of surviving connections.
• Structured pruning removes whole neurons (and every connection attached to them) whose average outgoing weight magnitude is low — the result is a smaller, dense network that runs faster on ordinary hardware.
• Accuracy tends to hold steady as sparsity rises, then drops sharply once pruning removes weights that were actually load-bearing — the "knee" of the curve.
• Real pruning pipelines usually fine-tune the surviving weights afterward to recover any lost accuracy; this simulation shows the structural effect only.

Controls

• Prune threshold — connections (or neurons, in structured mode) with magnitude below this percentile are pruned.
• Hidden layer width — resizes the two hidden layers and regenerates the network topology.
• Structured toggle — switch between per-weight and per-neuron pruning.
• Show pruned connections — keep pruned links visible as faint gray threads for comparison, or hide them entirely.
• ↺ Regenerate weights — reroll random weight magnitudes. Drag to rotate, scroll to zoom.

Did you know?

Some production pruning techniques remove over 90% of a network's weights while keeping accuracy within a percentage point of the original — a core trick for running large models on phones and edge devices.