Pulse Prune
Click the dim weights. Keep the bright ones.
How to play
A small neural network is drawn on the canvas: a few input neurons, some hidden neurons, a few output neurons, with every pair connected by a weight. Bright lines are high-magnitude weights. Faint lines are low-magnitude weights. Occasionally a blue pulse flows from an input through the live connections out to an output — that’s an inference running.
Your job: click the dim weights to remove them. Every dim cut scores +1. But if you click a bright weight, accuracy drops — because bright weights are doing real work. Reach 60% sparsity within 45 seconds while keeping accuracy above 50% to win.
The Systems Concept
Trained neural networks are full of small, near-zero weights that contribute almost nothing to the output. Han et al. (2015) showed that magnitude is a usable proxy for weight importance — you can cut the small ones, fine-tune briefly to recover, and substantially shrink the model. You’re doing that same decision by eye: telling the redundant weights apart from the load-bearing ones. The lottery-ticket hypothesis (Frankle & Carbin 2018) goes further and suggests the bright survivors, if trained from scratch, can match the full model’s accuracy.
Part of MLSysBook Playground. Found a bug? Report an issue.