Operads, Modules, and Differentiation in Machine Learning
The paper develops a concrete application of Kříž and May's one-marked-input modules to differentiable and stochastic programs. An operad encodes typed substitutions. Its algebra interprets their values, and the module interprets perturbations in one selected input.
For a smooth interpretation of a colored operad $P$, put $A_X=\mathbb R[X]$ and $M_X=\bigoplus_{x\in X} T_xX$. The free real linearization $\mathbb R[P]$ acts on $A$ by applying each operation to its basis inputs. The partial derivatives of the interpreted operation define the marked-slot actions on $M$. The chain rule supplies the module substitution axiom. The associated square-zero extension is the quotient of the free vector space on the total tangent bundle by the fiberwise affine relations.
The paper derives the reverse sweep by dualizing local derivative maps and accumulating covectors. For a shared parameter, the diagonal pulls the slot covectors back to their sum. A separate hyperdual construction computes mixed second derivatives. Exact and numerical worked examples cover a tied nonlinear layer and a Deep Sets pooling map.
For finite symmetry groups, the paper distinguishes ordinary orbit averaging in characteristic zero from stabilizer homology in modular derived coinvariants. For finite stochastic kernels, signed measures of mass zero form another operadic module. A tree bound estimates how local kernel discrepancies propagate in total variation under independent child evaluation.
This is a mathematical treatment with reproducible checks of the worked examples. The full manuscript is attached as a double-column PDF. The standalone UTF-8 LaTeX contains the complete text, figures and bibliography and is designated for search indexing after publication. The ZIP contains the organized source and Python verification script.
Authorship and assistance
The responsible submitter requests an anonymous byline. A language model assisted with drafting, typesetting, mathematical review, and revision.