LabRuns entirely in your browser. Nothing is sent to any server.
Switch off 90% of the brain
Magnitude pruning on the digit classifier: sort the weights by size and switch off the smallest. The weight matrix visibly goes dark pixel by pixel, and accuracy — measured live over 300 images — holds up to 80% pruning before falling off a cliff. Trained networks are enormously redundant.
The other way to shrink a model: fewer bitsMore experiments: