Blog¶
Plain-language write-ups of the research behind PolyWeave — each grounded in the actual library code, and adaptable to Medium / Towards Data Science.
| Post | Topic | Level |
|---|---|---|
| Generating a Network's Weights with a Hypernetwork | Build a hypernetwork with PolyWeave on CIFAR-10 — zero-shot weight generation, a faster warm start, and a cheap ensemble. | Beginner |
| Logic, in Products: Gates, Chaining, and Rule Induction | Fuzzy gates, XOR in one neuron, differentiable forward chaining, and interpretable rule induction with negation. | Beginner |
| How Linear Is a Transformer's Feed-Forward Block? | Measuring per-block linear recoverability; why it's learned, not architectural; and using it as a compression criterion. | Intermediate |
| When Does Multiplication Help a Neural Network? (coming soon) | The pi-scale recruitment diagnostic — what it measures, and an honest account of what it does and doesn't tell you. | Intermediate |
These accompany two arXiv papers
"How Linear Is a Transformer Feed-Forward Block?" and "When Does the Pi Branch Fire?". The posts summarise the ideas for a broader audience; the papers have the full experimental detail. Companion Jupyter notebooks are planned (see each post).