Skip to main content

Frequently asked questions

What is omnibias in one sentence?

A framework that computes the closed-form n-th derivative of an activation, σⁿ(z), for arbitrary n, with bit-stable accuracy and a single σ evaluation — exposed identically on PyTorch, JAX, and Keras 3.

Is it a replacement for autograd?

No — it is a targeted accelerator. Use autograd for general gradients; reach for omnibias when you need high-order or iterated derivatives (Laplacians, Hessians, Δᵏ) where nested autodiff is slow or memory-bound. See why omnibias.

How can a closed form be "faster" and "identical"?

Because it computes the same mathematical object a cheaper way: one more recurrence coefficient instead of another differentiation pass. The benchmarks report the max abs difference (≤ 1e-15) alongside every speedup. See complexity.

Which activations support high orders?

The smooth (Riccati) family — tanh, sigmoid, softplus, gaussian — at any order. relu/silu/gelu support only value and gradient. See the activation dictionary and stability matrix.

Why are results bit-identical across backends?

There is exactly one definition of the polynomial coefficients, in pure-Python omnibias-core; every backend imports it. See cross-backend parity.

Does it work with float32?

Yes, but the parity and certificate guarantees are float64. Use float64 for any correctness or rigor claim, and float32 only for throughput when your application tolerates it.

Is omnibias a PDE solver?

No. It accelerates the derivatives used in PDE residuals and can bound the residual of a trained surrogate over a box — but it does not prove existence, uniqueness, or regularity of true solutions. See scope boundaries.

What does a certificate actually prove?

That a specific enclosed quantity lies within the stated bounds, to outward-rounded rigor, and that the record was not tampered with. If theorem_prover_verified is true, a Lean kernel also checked the finite obligation. See provenance & certificates.

Does it need a GPU?

No. Everything runs on CPU; the smoke benchmark tier is CPU-friendly. An accelerator helps for large full-tier runs.

Can I use a pretrained relu/gelu backbone?

Yes — keep its activation and use op="identity". You only need a smooth activation where you actually take high-order derivatives.

How do I add a custom activation?

Register an ActivationSpec with its fast path and supported orders. See activation dictionary.

Where do I report a bug?

Open an issue at github.com/derivon-ai/omnibias/issues with a minimal reproduction in float64, the exact omnibias-* versions, and a seed. A float64 cross-backend mismatch or an enclosure that excludes a true value is always a bug.

For anything security-sensitive, do not open a public issue — use GitHub's private vulnerability reporting on the repository, or email info@derivon.ai with the subject line SECURITY:.

Which licence applies to me?

Depends on the package. 28 of the 42 distributions — the derivative tower, every backend, and everything built directly on them — are Apache-2.0, so closed-source and hosted use carry no obligation. The 14 certified-decision packages are AGPL-3.0-or-later or commercial. No Apache package ever pulls an AGPL one into your dependency tree. See Licensing.

How do I get commercial terms, or talk to a human?

Email info@derivon.ai — commercial licensing, support, partnerships, and security reports all route there. General technical questions are better as a GitHub issue so the answer is public.

See also