WeightScope is a real-time instrument for watching a neural network's weights evolve while it trains. Every parameter in the network is laid out along a Hilbert curve, a space-filling curve that keeps neighboring weights near each other on screen, so a 20-million-parameter model becomes a single coherent image that updates live as training runs.
The map is interactive. A paintable brush lets you select any region of weight space mid-training and freeze it (hard-pin the weights) or zero it (ablate the weights), then watch the effect on train and validation loss immediately. The captioned demo above shows a ResNet-18 training on CIFAR-100: it tours the three view modes (weight value, |Δw|, and |grad|), then freezes a patch of early layers, holds a transient Zero ablation across the middle of the network while the loss reacts, and finishes with permanent Sparsify stamps.
Most training-dynamics work studies networks through scalar summaries: loss curves, gradient norms, spectral statistics. WeightScope treats the full weight state as the object of study, making the spatial structure of learning visible: which regions of a network move, when they move, and how much damage the network absorbs when they are pinned or removed. It serves as lab equipment for experiments in weight-space criticality, graded ablation, and out-of-distribution robustness.