Watch LeWi train

Neurons move, connections form and break, and the wiring comes to align with the target.

Pick a target and press play. The colors are assigned after training; the network never sees them.

Training, step by step

A network with 32 neurons per hidden layer and four hidden layers learns a depth-4 tree over eight inputs. Each neuron takes the color of the target piece its activity follows most closely; unassigned neurons, in gray, follow none clearly.

Neuron positions

drag to rotate · scroll to zoom · hover for details · click to highlight a path · shift-click to trace connections

Edges Edge opacity

Target

Same colors as the neurons.

click a node to isolate its neurons · shift-click to add

Validation error (NMSE)

drag to scrub

Connection matrices

Soft gate g

Binary mask m

Effective weight W˜ = W g m

rows: receiving neurons · columns: sending neurons · side strips: each neuron's color · hover a cell for details, click to pin it

loading…
Jump to a rewiring:

Space plays or pauses; ←/→ steps (shift: ×10).

loading the run…

The final wiring

The trained network, one panel per layer transition plus the whole network. All panels turn together: drag to rotate, scroll to zoom, click a neuron to trace its connections.