Skip to main content
Open-core · PyTorch · JAX · Keras 3

The n-th derivative of an activation, in closed form

Numerically-stable closed-form n-th derivative framework. One sigma evaluation gives the whole derivative tower — bit-stable to machine precision, identical across backends, and certifiable where it matters.

pip install omnibias-torchpip install omnibias-jax

Commercial demos and partnerships: info@derivon.ai

199x
faster Laplacian
GPU · vs torch.func.hessian, D=240
108x
less memory
GPU · vs dense Hessian, D=240
480x
faster Δ³
GPU · vs nested forward-Laplacian
1e-15
agreement
float64, bit-for-bit

Headline figures are float64 on a single data-center GPU (H=256, B=4096). Every method agrees to ≤ 1e-15 — speedups are bit-for-bit. See benchmark methodology.

One primitive, a full operator surface

omnibias turns the closed-form activation derivative tower into every differential operator scientific ML needs.

Closed-form n-th derivative

Call σⁿ(z) for arbitrary n with one activation evaluation — polynomial recurrences from the Riccati / Eulerian / Hermite identities.

O(1) Laplacian in dimension

The closed-form Laplacian overhead is independent of input dimension D. Grow D 80x and the wall-clock barely moves.

Bit-identical backends

PyTorch, JAX, and Keras 3 import the same pure-Python coefficients, so every (activation, order) pair is float64-ULP-equal.

Certified register

Interval / Taylor-model arithmetic produces provable enclosures, sealed in tamper-evident certificates with a Lean-kernel gate.

Iterated operators stay flat

Poly-Laplacians Δᵏ for biharmonic and beyond cost one tower per order — no nested-autodiff blow-up or OOM.

Full operator surface

Gradient, divergence, curl, Hessian, Jacobian, Laplace-Beltrami, exterior derivative, and multivariate jets.

A 30-second tour

Drop a typed operator into your model and keep your training loop.

quickstart.py
import torch
from omnibias.torch import OMBU, OperatorBlock, cmbLinear

# OMBU: trainable K-bias operator, drop-in for an activation.
ombu = OMBU(num_channels=4, K=2, base="tanh")
out = ombu(torch.zeros(8, 4))

# Typed scalar operator (grad / laplacian / integral / identity).
laplacian = OperatorBlock(channels=8, op="laplacian", base="gaussian")

Full quickstart →

Who is omnibias for?

PINN buildersLaplacians and high-order PDE residuals without autograd-through-layers.
Quantum chemistsLocal kinetic energy for FermiNet / VMC, bit-identical to the default.
Optimization researchersClosed-form Hessian, Fisher, and KFAC factors for second-order methods.
Verification teamsCertified residual, robustness, and reachable-set enclosures.

Make high-order derivatives cheap

Start with the quickstart, then wire the closed-form path into your PINN, VMC, or optimization loop.

Or email info@derivon.ai