Why use omnibias?

Scientific residuals often ask for a derivative or contraction that a general-purpose differentiation graph computes indirectly. omnibias exposes mathematical representations chosen for those quantities.

• Activation identities: evaluate supported high-order derivatives using shared polynomial recurrences.
• Directional jets: carry derivatives along a few selected directions through a network.
• Mixed jets: represent all requested mixed partials, with an explicit combinatorial coefficient budget.
• Direct contractions: compute Laplacians without storing a full Hessian.
• Trainable decisions: use soft gates and structured relaxations inside the loss.

These are complementary tools. Parameter autodiff remains part of the learning process. Some workloads favor alternative approaches, including JAX Taylor-mode AD; consult the benchmark results.

Start with the quickstart and choose an API by the quantity your model actually needs.