Math that trains
omnibias provides forward spatial derivatives and trainable decisions for scientific machine learning. Supported activation identities, Taylor composition, and direct operator contractions replace nested high-order spatial differentiation while retaining gradients into model parameters.
Choose your first path
| You need | Start here |
|---|---|
| An activation derivative | Five-minute quickstart |
| A high-order trainable residual | Fourth-order tutorial |
| A Laplacian without a Hessian | Direct contraction |
| Trainable region boundaries | Soft regional model |
| Curvature-aware optimization | Gauss–Newton tutorial |
| A precise API contract | Package explorer |
What is public
The current workspace contains 16 packages. Core and backend releases are RC/beta; extension packages are explicitly marked alpha. The installation guide pins the versions used here. The ecosystem guide distinguishes primitives from separate application solvers.
What the claims mean
These representations remove a specific nested-graph mechanism on supported paths. Arithmetic still grows with derivative order, dimension, width, or the number of coefficients being requested. Accuracy is measured separately from runtime. A useful primitive does not establish convergence of every PDE or every training procedure.
Read scope and guarantees, then run a small example against an independent reference.