Notebooks

Two marimo notebooks walk the public API end-to-end using a single running example: a small classifier trained on the two-moons dataset. Both compute the sparse and exact partitions before, during, and after training, look at the partition layer by layer, and animate the full training run — then continue into inspecting a Partition, analysis/verification functions, a higher-dimensional example, and save/reload. The only difference between them is the plotting backend.

A third notebook asks a more specific question: do linear regions survive into feature-space? It projects hidden-layer activations to 2D with tSNE/UMAP and checks whether points from the same linear region cluster together, with an interactive architecture/layer picker.

These are static, pre-exported snapshots (marimo export html) rather than live notebooks — parx depends on Julia and PyTorch, neither of which run inside a browser sandbox, so a fully interactive in-browser version isn't possible here. To run any notebook live and edit it, clone the repo and run:

uv run notebooks/demo_plt.py            # matplotlib backend
uv run notebooks/demo_plotly.py         # plotly backend
uv run notebooks/feature_embedding.py   # feature-space embedding

Matplotlib backend

Static figures, no hover tooltips; the training animation is a plain FuncAnimation.

Plotly backend

Interactive figures with hover tooltips; the training animation has a play/pause button and slider.

Feature-space embedding

Do linear regions survive into feature-space? Projects hidden-layer activations down to 2D with tSNE/UMAP and colors points by which linear region they belong to, across a selectable architecture and extraction layer.