Interactive method visualization

A frozen network finds its own route.

Double-Scoring trains scores around fixed random weights, allowing a sparse subnetwork to emerge without training the weights themselves.

Illustrative scikit-learn digits run Not one of the paper’s FashionMNIST benchmark experiments
Epoch 0

Play the genuine saved checkpoints or scrub through them yourself; use Resume to continue from the checkpoint you choose. Choosing a digit probes the current mask; it does not change training.

Network view

Scores compete. A ticket lights up.

Loading graph…

Thickness = |weight| Opacity = score margin above cutoff Masked links hidden Positive weight Negative weight Glow = newly selected Hidden node size = activation Output halo area = softmax output

Double-score competition

Selected-slot allocation

Each layer keeps the top 18% of scores across its real connections and an equally large pool of zero-weight dummy coordinates. A selected real slot becomes a live edge; a selected dummy slot does not. Each bar shows the composition of that fixed selected quota—not the size of the two candidate pools.

Try another input

Follow ten examples through the same ticket.