\documentclass[10pt,letterpaper,twocolumn]{article}
\usepackage[T1]{fontenc}
\usepackage{lmodern}
\usepackage[
  letterpaper,left=.72in,right=.72in,top=.75in,bottom=.73in,
  headheight=14pt,headsep=14pt,footskip=26pt
]{geometry}
\usepackage{microtype}
\usepackage{mathtools,amssymb,amsthm}
\usepackage{xcolor}
\usepackage{tikz-cd}
\usepackage{titlesec}
\usepackage{fancyhdr}
\usepackage{caption}
\usepackage{booktabs,balance}
\usepackage{xurl}
\usepackage{hyperref}
\usepackage{bookmark}

\definecolor{paperink}{HTML}{202824}
\definecolor{paperaccent}{HTML}{145A51}
\definecolor{papermuted}{HTML}{51655F}
\definecolor{paperrule}{HTML}{B6CBC4}
\hypersetup{
  pdftitle={Topos Actions in Machine Learning and Artificial Intelligence},
  pdfauthor={Anonymous},
  pdfsubject={Prior work and models for local symmetry},
  colorlinks=true,linkcolor=paperaccent,citecolor=paperaccent,urlcolor=paperaccent
}
\AtBeginDocument{\color{paperink}}
\setlength{\columnsep}{.29in}
\setlength{\parindent}{1em}
\setlength{\parskip}{1pt plus .5pt}
\setlength{\emergencystretch}{1.7em}
\linespread{1.035}
\allowdisplaybreaks[1]
\raggedbottom
\clubpenalty=1000
\widowpenalty=1000

\titleformat{\section}{\large\bfseries\raggedright}
  {\color{paperaccent}\thesection}{.55em}{}
\titlespacing*{\section}{0pt}{14pt plus 3pt minus 2pt}{5pt}
\titleformat{\subsection}{\normalsize\bfseries\raggedright}
  {\color{paperaccent}\thesubsection}{.55em}{}
\titlespacing*{\subsection}{0pt}{10pt plus 2pt minus 1pt}{4pt}
\captionsetup{font=small,labelfont={bf,color=paperaccent}}

\pagestyle{fancy}
\fancyhf{}
\fancyhead[L]{\footnotesize\sffamily\color{papermuted}Topos actions in machine learning and artificial intelligence}
\fancyfoot[R]{\footnotesize\sffamily\color{papermuted}\thepage}
\renewcommand{\headrulewidth}{0pt}
\fancypagestyle{plain}{%
  \fancyhf{}%
  \fancyfoot[R]{\footnotesize\sffamily\color{papermuted}\thepage}%
  \renewcommand{\headrulewidth}{0pt}%
}

\theoremstyle{plain}
\newtheorem{theorem}{Theorem}[section]
\newtheorem{proposition}[theorem]{Proposition}
\newtheorem{lemma}[theorem]{Lemma}
\newtheorem{corollary}[theorem]{Corollary}
\theoremstyle{definition}
\newtheorem{definition}[theorem]{Definition}
\newtheorem{example}[theorem]{Example}
\theoremstyle{remark}
\newtheorem{remark}[theorem]{Remark}
\DeclareMathOperator{\Map}{Map}
\DeclareMathOperator{\Tot}{Tot}
\DeclareMathOperator{\Fun}{Fun}
\DeclareMathOperator{\Hom}{Hom}
\DeclareMathOperator{\grad}{grad}
\DeclareMathOperator{\Ext}{Ext}
\DeclareMathOperator{\Res}{Res}
\DeclareMathOperator{\Kl}{Kl}
\DeclareMathOperator{\id}{id}
\DeclareMathOperator{\rank}{rank}
\newcommand{\E}{\mathcal E}
\newcommand{\Spc}{\mathcal S}
\newcommand{\R}{\mathbb R}
\newcommand{\C}{\mathbb C}
\newcommand{\Z}{\mathbb Z}
\newcommand{\overB}{\E_{/B}}
\tikzcdset{every label/.append style={font=\small,/tikz/auto=false,description}}
\begin{document}
\twocolumn[
\begin{@twocolumnfalse}
\noindent\textcolor{paperaccent}{\rule{\textwidth}{1pt}}\par
\vspace{11pt}
{\noindent\fontsize{25}{30}\selectfont\bfseries
Topos Actions in Machine Learning\\
and Artificial Intelligence\par}
\vspace{8pt}
{\noindent\large\color{papermuted}
Prior work and models for local symmetry\par}
\vspace{10pt}
{\noindent\small November 2026\par}
\vspace{9pt}
\noindent\begin{minipage}{\textwidth}\small
\textbf{Abstract.}
We review uses of topoi, group actions, and descent in machine learning
and artificial intelligence. Existing work models neural networks by
presheaves and stacks, learns automata with nominal symmetries, and
builds architectures from cellular sheaves or groupoid-equivariant
kernels. The literature identified here gives little direct treatment
of learning with a monoidal Grothendieck topos acting by geometric
morphisms. We develop two applications in that setting and its
internal counterpart. A central action gives a quotient-type criterion
for nuisance-invariant layers and distinguishes quotient features from
linear averaging. A nonconstant group sheaf over the circle yields
learnable filters whose coefficient bundle has a M\"obius summand.
Descent determines which parameter choices and update rules give
consistent models across contexts. The results are structural and
are illustrated by exact learning examples.
\par\medskip
\noindent\textbf{Keywords.}\enspace
Topos actions, equivariant learning, descent, sheaf neural networks,
nominal automata, nuisance invariance.
\end{minipage}
\vspace{17pt}
\end{@twocolumnfalse}
]
\thispagestyle{plain}

\section{Which notion of action enters learning?}
\label{sec:introduction}

Equivariance in machine learning usually begins with one group,
its representations on data and features, and a trainable map
that respects those representations. Three developments motivate
a topos description. A local model may depend on a varying
context. The observations and the symmetry groups themselves
can have nontrivial transition maps. An action can also relate
different systems of observations by geometric reindexing.
Keeping those distinctions explicit lets one say which facts
are architectural and which are consequences of descent.

In the geometric definition of \cite{topos-actions},
the parameter $\mathcal T$ is a monoidal Grothendieck
topos. Its action on $\mathcal X$ is a geometric morphism
$a:\mathcal T\times\mathcal X\to\mathcal X$ with coherent
multiplication and unit. Its inverse image
$a^*$ pulls an observation on $\mathcal X$ back to an observation
on the product with the symmetry parameter. A second construction
is a weak action of an ordinary group by geometric
autoequivalences. It is realized by taking $\mathcal T$ to be
the sheaf topos of the discrete locale of group elements.
Finally, a group object $G$ in a fixed topos $\mathcal E$ acts
on objects of that topos. The topos $B_{\mathcal E}G$ of internal
$G$-objects classifies compatible families with those symmetries.
A locally constant sheaf of groups illustrates why its group of
global sections may be insufficient to specify a model.

For a finite group in sets this distinction is visible even
before learning. The parameter topos $\mathrm{Set}^{G}$ indexed
by the underlying discrete set of elements has points indexed
by $G$. The topos of $G$-sets, denoted $\mathbf B G$, classifies
$G$-torsors. Its points form a groupoid with one isomorphism
class and automorphism group $G$. Group laws on the set of
elements induce geometric operations on the parameter topos.
A Cartesian monoidal law on $\mathbf B G$ requires $G$ to be
abelian, as proved in \cite{topos-actions}. Below we use an
actual action of $\mathbf B A$ for abelian $A$ and compute
what it says about learning a nuisance-invariant readout.

\subsection{External symmetry and spaces of layers}

Let an ordinary group $H$ act by geometric autoequivalences on
a Grothendieck topos $\mathcal E$. Write $T_h$ for the inverse
image indexed by the inverse element, so that
$T_hT_k\simeq T_{hk}$. An equivariant object $X$ has structure
$\theta_{h,X}:T_h X\simeq X$. The equivariant topos
$\mathcal E^H$ is the geometric quotient. For equivariant
objects $X,Y$, the underlying exponential $Y^X$ acquires
the conjugation action. The equivalence
$T_h(Y^X)\simeq(T_hY)^{T_hX}$ is followed by precomposition
with $\theta_{h,X}^{-1}$ and postcomposition with $\theta_{h,Y}$.
The two coherence laws follow by composing these maps on
either side of an evaluation. Consequently a layer
$P\times X\to Y$ is equivariant precisely when its transpose
$P\to Y^X$ is equivariant. The usual equation
$F(hp,hx)=hF(p,x)$ is its case in sets. It is an equation
about the family. At a fixed parameter the equation applies
to its stabilizer.

This elementary result makes the applications testable.
An architecture specifies a parameter object, a realization
map, and which symmetries it respects. A loss and a learning
rule require their own compatibility with those data.
Sections~\ref{sec:direct}--\ref{sec:categorical-ml} review
relevant work. The later sections develop descent for layer
evaluation and compute two specializations where the
symmetry changes what can be learned or represented.
\par
\section{Topos-theoretic precedents in learning}
\label{sec:direct}

The closest precedents use three different levels of structure.
Some place the states and maps of a learner in a topos of equivariant
objects. Others use a classifying topos to organize possible
interpretations and changes of context. A literal action in the
sense considered here gives a monoidal Grothendieck
topos \(\mathcal T\) a geometric action
\(\mathcal T\times\mathcal X\to\mathcal X\) on another topos. One
can test this last formulation by looking for the parameter topos,
the geometric action map, and the coherent multiplication and
unit comparisons. The existing overview of Jia, Peng, Yang, and
Chen \cite{jia-topos-survey} collects examples of topos-based
learning in the broader sense. We examined its references,
then checked primary sources for geometric actions, classifying
topoi, equivariant sheaves, and isotropy in relation to learning.
The sources give a basis for a focused survey at the first
two levels. They leave the design of learning procedures using
literal monoidal-topos actions or general localic isotropy as a
separate question.

\subsection{Neural networks in equivariant topoi}

A direct precedent relating topoi and neural networks is
Belfiore and Bennequin's preprint \cite[\S2.1--2.2]{belfiore-bennequin}.
A network has an indexing category \(C\) that records its
architecture. In their constant-symmetry example one replaces
set-valued presheaves by presheaves taking values in
\(\mathbf{B}G\), the topos of left \(G\)-sets for a discrete
group \(G\). A feature diagram is then
\[
  F:C^{\mathrm{op}}\longrightarrow \mathbf{B}G .
\]
Arrows of \(C\) induce restriction maps on feature data and
the maps of \(\mathbf{B}G\) require equivariance of each
transition. Curried evaluation gives
\[
\begin{aligned}
 \operatorname{Fun}(C^{\mathrm{op}},\mathbf{B}G)
   &\simeq \operatorname{Fun}(C^{\mathrm{op}}\times BG,\mathbf{Set})\\
   &\simeq \widehat{C\times(BG)^{\mathrm{op}}}.
\end{aligned}
\]
The last term is the presheaf topos on \(C\times (BG)^{\mathrm{op}}\). Here \(BG\)
is the one-object category with arrow group \(G\). In
equivariantization terminology the first expression is the
topos of equivariant objects for the trivial external action of
\(G\) on \(\widehat C\). This is a concrete use of the group
classifying topos as a target for feature data. The geometric
theory of arbitrary \(G\)-actions has a classifier of its own.
Its points describe \(G\)-sets, whereas points of \(\mathbf{B}G\)
describe \(G\)-torsors \cite[slide 22]{lafforgue-eth-ai}.

Belfiore and Bennequin also let the categories in the
invariance structure vary over the architecture and use
a fibration whose fibers carry those categories. Their relative
action is given by contravariant functors between fibers
commuting with reindexing \cite[Eq.~(2.1)]{belfiore-bennequin}.
This specifies compatible diagrams of behaviors in the
fibers. A geometric action of a parameter topos would further
specify a multiplication on the parameter topos and a coherent
geometric morphism out of a product. The preprint proposes that
invariance explains aspects of extrapolation. The topos
description supplies a type of invariance constraint.
Statistical generalization still requires an input distribution,
a loss, and hypotheses about a training procedure.

A related result gives a topos from the selected interface of
a learner. Spivak \cite{spivak-learners-languages} represents
parametrized learners as generalized Moore machines using
polynomial interfaces. For a fixed polynomial \(p\), the
category of \(p\)-coalgebras forms a topos. Its internal language
can express propositions about the dynamics, with gradient
descent as a worked case. This makes the topos a setting for
specifying learner state and behavior. The monoidal structure
used to build the interfaces belongs to the category of
polynomials. The more recent survey also discusses
\emph{The Topos of Transformer Networks}. The canonical record
for that preprint is marked withdrawn
\cite{villani-transformers-withdrawn}.

\subsection{Relative classifying topoi for symbolic AI}

Lafforgue \cite[slides 20--25]{lafforgue-eth-ai} proposes
organizing a data language by a geometric theory \(T\).
A map from a test topos to its classifying topos classifies a
\(T\)-model in that test topos. Taking the test topos to vary
therefore models families of interpretations and their
change of base. His proposed networks of topoi attach
geometric object categories and classifiers to different
processing stages. A transformation natural in whole
families is a candidate for a geometric morphism between
classifiers. This provides a way to formulate the transfer of
structured representations. The talk gives an architecture
proposal rather than an optimization or accuracy result.
Caramello's course on relative topoi \cite{caramello-raven}
makes the symbolic application more concrete. It writes
invariance constraints for Raven matrices in a geometric
language, including an example where an arithmetic rule on
positions is expressed by a formula indexed by the graph of
that rule. Restricted vocabularies and specified relations
make candidate interpretations comparable by proof. This is
a formal model of a class of reasoning problems and a distinct
application from training an equivariant numerical predictor.

\subsection{An algorithmic precedent via nominal sets}

Active automata learning gives an implemented example of
learning in a symmetry-sensitive setting. The nominal topos
is the category of discrete sets with continuous action of the
permutation group of countably many names, equipped with
the topology of pointwise convergence. Continuity of an orbit
map means that an element has a finite support: its stabilizer
contains the subgroup fixing some finite set of names
pointwise. The set of name pairs, for example, has two orbits,
equal and unequal pairs. Such orbit-finite structure permits
a finite orbit description even for machines with infinitely
many individually named inputs. Iwaniack
\cite{iwaniack-automata-toposes} uses the nominal topos as
an instance of topos-enriched automata and derives
Myhill--Nerode results there.

Moerman and collaborators \cite{moerman-nominal-learning}
adapt Angluin's query algorithm to nominal automata, including
preliminary experiments with a Haskell implementation.
The teacher answers queries about acceptance and proposes
counterexamples to incorrect hypotheses. Equivariance lets
sets of examples and hypotheses be handled by their orbit
structure. Barb\'e \cite[Thm.~4.26]{barbe-topos-automata}
generalizes this formulation in a research report.
Over an atomic topos with countable coproducts, an
alphabet and target state object each having finitely
many atom components, and an appropriate teacher,
\(\mathrm{TopL}^{*}\) terminates with a minimal automaton.
This conclusion concerns exact learning in the stated
query model. An effective implementation still depends on
the ability to manipulate the orbit or atom presentations
and to supply the teacher. These are meaningful operational
assumptions when trying to infer a structured protocol model
whose states depend on relationships among names.
\par
\section{Sheaf learning and local equivariance}
\label{sec:sheaves}

Two established lines of work supply concrete models for
applying topos-action ideas. Spectral sheaf learning builds
features that satisfy incidence constraints. Equivariant
network design builds maps that commute with transports.
Their intersection can be described exactly for invertible
transports on a graph.

\subsection{From sheaf Laplacians to learned incidence maps}

Hansen and Ghrist develop the spectral theory of cellular
sheaves, including Hodge Laplacians and their relation to
sheaf cohomology \cite{hansen-ghrist-sheaves}. Hansen and
Gebhart use the sheaf Laplacian in graph neural layers
\cite{hansen-gebhart-snn}. Bodnar et al.\ learn the
incidence maps from data, study how sheaf geometry changes
diffusion and oversmoothing, and compare their models on
graphs where adjacent vertices can have different labels \cite{bodnar-nsd}. An earlier direction
in graph learning uses edge-conditioned filters
\cite{simonovsky-ecc}. Edge attributes there determine
message weights. In a sheaf layer, maps from the two
endpoints into a common stalk determine which local
features agree and how disagreement contributes to the
diffusion.

For a finite loopless graph $\Gamma$, write $V_v$ for a vertex
stalk and $V_e$ for an edge stalk, with maps
$r_{ev}:V_v\to V_e$ at each incidence. Give the finite
dimensional stalks inner products and orient each edge
$e:u\to v$ for purposes of signs. Set
\[
  (\delta x)_e=r_{ev}x_v-r_{eu}x_u,\qquad
  L=\delta^*\delta .
\]
The Laplacian satisfies
$\langle x,Lx\rangle=\|\delta x\|^2$.
Thus $\ker L=\ker\delta$ is the vector space of global
sections, and $e^{-tL}$ converges to orthogonal projection
onto that space as $t\to\infty$. This last statement is
for the fixed unnormalized operator on a finite graph.
It follows by diagonalizing the positive semidefinite
operator and does not describe an arbitrary learned
nonlinear diffusion trajectory.

Suppose each incidence map is invertible. Then the edge
transport from $u$ to $v$ is
$T_e=r_{ev}^{-1}r_{eu}$, and the reverse edge transports
by $T_e^{-1}$. These maps extend to a representation of
the fundamental groupoid of the graph, regarded as a
one-dimensional complex. Indeed, the incidence category has
the vertices and edges as objects and arrows from a vertex to
its incident edges. Localizing by those arrows gives the
fundamental groupoid of the barycentric subdivision. By the
universal property of localization, precisely the diagrams with
invertible incidence maps extend. Restricting to the original
vertices gives the edge transports $T_e$. Global sections correspond to
parallel families $x_v=T_e x_u$. On each connected
component choose a base vertex $b$. Transport of $x_b$
determines the whole family, and it is independent of
the chosen paths precisely when $x_b$ is fixed by every
based loop transport. Hence the global sections on a
connected graph identify with the holonomy-invariant
subspace of $V_b$. The assertion concerns graph paths.
If two-cells are added and their boundary loops are
imposed as relations, their transports must satisfy the
additional flatness conditions.

A rank-one example is a triangle with two positive unit
transports and one negative unit transport. The product
around the cycle is $-1$, so the only real global section
is zero. By comparison, using three positive unit
transports gives the line of constant sections. The
underlying incidence graph is unchanged. This separates
the effect of the transport from that of connectivity.
When the incidence maps are isometries, each residual
has norm $\|x_v-T_e x_u\|$. In that case the operator is
a discrete connection Laplacian in the sense relevant
to vector diffusion maps \cite{singer-wu-vdm}. For orthogonal
changes of stalk coordinates, let $Q_0$ and $Q_1$ be the block
diagonal changes at vertices and edges, respectively. The new
incidence matrices are $r'_{ev}=Q_e r_{ev}Q_v^{-1}$, giving
$\delta'=Q_1\delta Q_0^{-1}$ and $L'=Q_0 L Q_0^{-1}$.
Thus the fixed diffusion is compatible with these changes of
frames even when the incidence maps are noninvertible.

The same argument constrains learnable linear parameters.
For two representations $V,W$ of the graph path
groupoid, a family of local linear maps
$A_v:V_v\to W_v$ is compatible with transport exactly
when
\[
  A_v T_e^V = T_e^W A_u,\qquad e:u\to v .
\]
Equivalently, $A$ is a parallel section of the bundle of
linear maps, with transport
$A\mapsto T_e^W A(T_e^V)^{-1}$. On a connected graph
its parameter space is therefore the space of
intertwiners of the two holonomy representations at
one vertex. For sign and trivial line holonomies
around the same cycle, a scalar parameter is forced
to vanish. For two copies of sign holonomy, any scalar
at the chosen vertex extends. Weight tying here is
dictated by representation type as well as by
the connectivity of the data. The invertibility
hypothesis matters: ordinary cellular sheaves also
permit noninvertible incidence maps and unequal stalk
dimensions, for which the incidence equation still
makes sense and there need not be transport along
groupoid arrows.

\subsection{From global groups to local bisections}

Global group actions determine the shared weights in
group-equivariant convolutions \cite{cohen-welling}.
Cohen, Geiger, and Weiler characterize equivariant
linear maps between fields on homogeneous spaces by
their convolution kernels \cite{cohen-geiger-weiler}.
On meshes, de Haan et al.\ use parallel transport to
define orientation-sensitive convolutions that respect
changes of local frame \cite{dehaan-mesh}.

For a finite data set carrying a group action, let
$K_{x,y}:V_y\to W_x$ be the blocks of a linear
layer and let $\rho_V(g;y)$ and $\rho_W(g;x)$ be the
feature transports. Equivariance is equivalent to
\[
 K_{gx,gy}
 =\rho_W(g;x)K_{x,y}\rho_V(g;y)^{-1}.
\]
It follows by comparing coefficients of a feature
supported at $y$. Each diagonal pair orbit has one
free block in the intertwiner space of its joint
stabilizer, since two transports from a representative
to the same pair differ by a stabilizer element.
Conversely an intertwiner at each representative
propagates to a well-defined equivariant kernel.

Ibort, Jimenez-Vazquez, and Perez-Pardo extend this type
of kernel analysis to groupoid representations with a
chosen pseudogroup of local bisections, measure, and
input and output bundles. Their preprint treats bounded
domains, where a geometric transformation can be
defined on part of the data space and pair types can
depend on the boundary \cite{ibort-groupoid-cnn}. The
choice of which bisections to act by and the integration
and support conditions enter the analytical layer
construction. In the groupoid descent picture, the
representation bundles supply transport data on the
features and on the kernel's source and target
variables. The kernel then acts on spaces of sections
and can mix distinct sites. A morphism of feature
representations over the identity base, such as $A$
above, acts fiberwise. This distinction separates useful
precedents for learning with groupoid descent from
studies of a monoidal Grothendieck topos acting on
another topos by geometric morphisms.
\par
\section{Compositional training and inference}
\label{sec:categorical-ml}

The categorical literature on training and probability supplies two
ingredients for learning in a setting with topos actions. Parametrized
maps provide a compositional description of local models, and
stochastic kernels provide a description of local inference.
To obtain a model over a groupoid of domains or in an action topos,
one supplies additional transport for the parameters, inputs, and
outputs. Compatibility is a condition on both the local maps and
the operations applied to them.

\subsection{Reverse differentiation and descent of updates}

Fong, Spivak, and Tuy\'eras make a learner into a compositional
morphism with a forward evaluation, a parameter update, and a
backward request. Under their choices of error function and fixed
learning rate, gradient descent gives a monoidal functor from
parametrized functions to learners \cite{fong-backprop}.
Cruttwell et al.\ separate the optimizer and loss from the architecture
using parametrized maps, lenses, and reverse derivative categories
\cite{cruttwell-gradient}. These constructions address how a training
rule composes along a network. To use them with an action on a topos,
the relevant categories and rules must also support transport along
changes of domain. The transport condition is what allows the
compositional training rule to respect the chosen action.

For example, in smooth models the cotangent pullback of a loss
covector is canonical. Given layers \(f:X\to Y\) and \(g:Y\to Z\),
the cotangent maps satisfy
\[
 D(g\circ f)_x^*=(Df_x)^*(Dg_{f(x)})^* .
\]
The same identity applied to a commutative square of local realizations
describes the transformation of the reverse pass. Converting a
cotangent vector into a parameter update adds geometric data. In
particular, a matrix transpose relative to a Euclidean metric in one
coordinate system implements a different conversion when a
nonisometric change of coordinates is made and the same Euclidean
formula is used in the new coordinates.

An explicit sufficient condition for a global gradient is available
in the smooth setting. Suppose a groupoid presents a parameter model
by local manifolds \(P_i\) with transition diffeomorphisms
\(\phi_{ij}\) on the corresponding overlaps. Equip the local
manifolds with losses \(\ell_i\) and positive-definite metrics \(g_i\)
satisfying
\[
   \ell_i=\ell_j\circ\phi_{ij},\qquad
   g_i=\phi_{ij}^{*}g_j .
\]
Then their gradients are compatible on the overlaps:
\[
  D\phi_{ij}(\operatorname{grad}_{g_i}\ell_i)
     =(\operatorname{grad}_{g_j}\ell_j)\circ\phi_{ij}.
\]
Indeed, pairing the left-hand side with \(D\phi_{ij}(v)\) in \(g_j\)
gives \(g_i(\operatorname{grad}_{g_i}\ell_i,v)=d\ell_i(v)\).
Differentiating the loss identity gives the same pairing for the
right-hand side. Surjectivity of the differential of the transition
and nondegeneracy of the metric give equality. Local gradient flows
therefore respect the transitions wherever defined, by uniqueness of
solutions. A discrete training rule requires its own compatibility.
For instance, a step made using the Riemannian exponential inherits
the compatibility on the domain where it is defined, because the
transitions are local isometries.

Natural gradient supplies a relevant source of metrics
\cite{amari-natural}. For a smooth family of strictly positive
probability laws \(p_\theta(y)\) on a finite outcome set, its Fisher
form is
\[
 I_\theta(v,w)
  =\sum_y \frac{d p_\theta(y)(v)\,d p_\theta(y)(w)}{p_\theta(y)} .
\]
If a transition reparametrizes the family and permutes the outcome
labels, the form transforms by metric pullback, by the chain rule
and reindexing the sum. When the likelihood map has injective
differential this form is positive definite and the descent argument
applies. In overparametrized models its degeneracy can require
passing to identifiable directions. Invariance of the resulting
natural gradient depends on the chosen model or pushforward
\cite{vanoostrum-natural}.

In a supervised problem, compatibility of the losses can itself
be tested. Let \(\mu_i\) be a finite distribution on local labelled
examples, let \(F_i\) be the prediction map, and let \(\lambda_i\)
be the example loss. The risk
\[
 \ell_i(\theta)=\sum_{x,y}\mu_i(x,y)
                 \lambda_i(F_i(\theta,x),y)
\]
is preserved if the transition transports the example distribution,
intertwines the predictions, and preserves the example loss.
The claim follows by substituting the transported examples in the
sum. This gives an intrinsic risk for data described in several
frames. Distinct datasets require explicitly specifying the risk
to be shared across frames.

\subsection{Probabilistic models under transport}

Markov categories give an abstract semantics for stochastic kernels,
conditional independence, and disintegration \cite{fritz-markov}.
Quasi-Borel spaces supply cartesian closed structure with probability
distributions, permitting higher-order probabilistic programs such
as random functions \cite{heunen-qbs}. Bayesian lenses describe
composition of prior-dependent inversions when the corresponding
conditionals exist \cite{smithe-bayes}. These frameworks address
probabilistic modeling and inference. Applying them to a data-domain
groupoid requires specifying what a domain change does to outcomes,
states, and probability laws.

In the finite case the extra condition can be checked directly.
On domains \(i,j\), take state and observation sets \(P_i,Y_i\) and
\(P_j,Y_j\), normalized priors \(q_i,q_j\), and likelihood kernels
\(K_i,K_j\). A transition consists of bijections \(u:P_i\to P_j\)
and \(v:Y_i\to Y_j\). If \(q_j(u\theta)=q_i(\theta)\) and
\(K_j(vy\mid u\theta)=K_i(y\mid\theta)\), then Bayes' rule commutes
with the transition whenever the observed evidence has positive
probability:
\[
  q_j(u\theta\mid vy)
  =\frac{K_i(y\mid\theta)q_i(\theta)}
         {\sum_{\theta'}K_i(y\mid\theta')q_i(\theta')}
  =q_i(\theta\mid y).
\]
The denominator follows by reindexing the evidence sum along \(u\).
Thus coherent transitions for priors and likelihoods induce coherent
transitions for posteriors wherever these are defined. The positivity
hypothesis matters when specifying pointwise inference. More general
disintegrations may be determined only almost surely and require
care about which version is transported.

Compatibility under invertible transports should also be distinguished
from consistency under marginalization. Abramsky and Brandenburger's
contextuality framework studies when families of distributions on
overlapping measurement contexts extend to a global distribution
\cite{abramsky-contextuality}. For an elementary illustration, place
an equal mixture of \((0,1)\) and \((1,0)\) on each of the pairs
\(AB,BC,CA\) of three binary variables. Their singleton marginals
agree. Any joint distribution with these marginals would impose
all three inequalities with probability one, but three binary
variables cannot be pairwise unequal. In multi-view learning this
example identifies a concrete modeling distinction: pairwise calibration
can hold even when there is no joint law realizing the local models.
A useful inference invariant here is existence of a joint law,
measured separately from agreement of the observed marginals.
\par
\section{Descent of equivariant learning data}
\label{sec:semantic-descent}

An equivariant layer may be specified in local coordinates
that change with the observation context. The action topos
gives a category in which to compare these specifications.
Let \(\mathcal E\) be a Grothendieck topos, \(B\) an object,
and \(G_B\) an internal group of \(\mathcal D=\mathcal E/B\).
For example, \(G_B\) can be pulled back from a group on
\(\mathcal E\), or can itself vary over \(B\).
Use \(\mathcal C=B_{\mathcal D}G_B\), the category of
left \(G_B\)-objects in \(\mathcal D\). It is again a
Grothendieck topos \cite{topos-actions}.
Its products are computed over \(B\) with the diagonal
action.

A parametrized layer is a morphism
\(h:P\times X\to Y\) in \(\mathcal C\), where \(P\) is the
parameter object. In internal notation the condition is
\[
 h(gp,gx)=g h(p,x).
\]
A fixed parameter accordingly yields an equivariant
predictor for the full group when that parameter is fixed
by the full action. More generally the condition on the
specialized predictor uses the stabilizer of the parameter.
Composition of two layers uses the product of their parameter
objects, with the same diagonal action. Reusing parameters
across layers amounts to inserting an equivariant map into
that product.

\subsection{Predictions, losses, and updates}

Supervision is an object \(D\) equipped with input and target
maps \(\xi:D\to X\) and \(\tau:D\to Y\). Let \(R\) be a score
object from \(\mathcal D\) with trivial \(G_B\)-action, and
let \(c:Y\times Y\to R\) be an equivariant comparison.
The induced loss is
\[
 \ell:P\times D\to R,\qquad
 \ell(p,d)=c(h(p,\xi(d)),\tau(d)).
\]
Thus \(\ell(gp,gd)=\ell(p,d)\). In \(\mathcal D\) it factors
through the joint orbit object \((P\times D)/G_B\), using
the coequalizer of the action and projection. Passing to the
product of the two orbit objects can lose values of the loss.
For example, in sets take \(P=D=\{1,-1\}\) with the sign
action by \(C_2\), \(X=D\), and \(Y=\{1,-1\}\) with
trivial action.
The layer \(h(p,x)=px\), constant target \(1\), and comparison
\(c(a,b)=(1-ab)/2\) give two loss values on two joint orbits.
The product of the separate orbit objects is a singleton.
An objective \(P\to R\) obtained by averaging the loss
uses a data distribution and an integration operation as
additional data. For a finite data set in sets with
real scores and probability weights \(\mu(d)\), the risk
is \(\sum_{d}\mu(d)\ell(p,d)\). If
\(\mu(gd)=\mu(d)\), simultaneous invariance of the
loss and reindexing the sum show that this risk is
invariant on \(P\).

An update is separate data \(u:P\times D\to P\) in
\(\mathcal C\). Equivariance now says
\(u(gp,gd)=g\,u(p,d)\). With the same invariant finite
sampling law, the law of the next parameter at \(gp\)
is the pushforward along \(g\) of the corresponding law
at \(p\). This follows by replacing the random sample
\(d\) by \(gd\), which has the same law.
For a parameter change \(r:P'\to P\),
the predictor reparametrizes by precomposition. An update
\(u'\) on \(P'\) is compatible if it satisfies
\[
 r\circ u'=u\circ(r\times\mathrm{id}_D).
\]
An isomorphism \(r\) transports the update uniquely by this
equation. The equation is an additional requirement for a
general \(r\), where precomposition alone leaves the result
of the old update in \(P\). For example, the update
\(p\mapsto p+1\) on \(\mathbb R\) has no compatible lift
to the single parameter value \(0\).

\subsection{A local-to-global criterion}

\begin{theorem}[Descent of layers and updates]
Let \(U\to 1\) be an epimorphism in \(\mathcal C\).
Suppose the objects \(P,X,Y,D\) are given in
\(\mathcal C/U\) with descent isomorphisms satisfying the
cocycle on \(U\times U\times U\). Give the restriction
of the trivial-action score object \(R\) its canonical
descent isomorphism. The local layer, input and target maps,
comparison, and update descend precisely when each map
commutes with the specified descent isomorphisms. The
global diagram is unique up to a unique compatible
isomorphism. The induced local loss descends through
the same comparison maps.
\end{theorem}

\begin{proof}
In a Grothendieck topos an epimorphism is effective
descent for objects. To recall the construction, the
transport on \(A\to U\) over \(U\times U\) identifies
\(a\) above \(v\) with its transported element above \(w\).
The unit, inversion, and cocycle define an internal
equivalence relation on \(A\). Its effective quotient
\(Q\) exists. The map \(A\to Q\times U\) is monic
because each transport within one fiber is the identity.
It is epic because locally a representative of each
class can be transported to the specified fiber.
Hence it is an isomorphism. This is the usual descent
calculation using exactness of a topos
\cite{maclane-moerdijk}. A map preserving the transports
preserves the relations and descends by the quotient
property. Pullback along \(U\to1\) is conservative and
faithful, so it detects the required diagrams and their
equations. It also preserves products. Consequently the
local composite defining the loss is the restriction of
the global one.
\end{proof}

A useful cover is the regular \(G_B\)-object
\(G_B\to1\) in \(\mathcal C\). The slice
\(\mathcal C/G_B\) is equivalent to \(\mathcal D\): take
the fiber of an equivariant object over the identity.
The inverse sends \(A\) to \(G_B\times A\) over \(G_B\)
with left translation on the first factor. For an
element \(a\) of an equivariant object \(q:Q\to G_B\),
the inverse to this product presentation is
\((q(a),q(a)^{-1}a)\). Thus ordinary local data
over the regular cover can encode equivariant global
data, with the action recovered from the overlap
transports. To descend a local parameter choice, treat
it as a morphism from the terminal object of
\(\mathcal C/U\). The same compatibility condition
then tests whether it gives a global section of \(P\).

\subsection{Coordinate compatibility of gradient updates}

The update equation also applies to the Riemannian rules
of Section~\ref{sec:categorical-ml}. If the parameter
transition carries the parameter metric and loss, it
transports the gradient and the geodesics used for a
discrete step \cite{lee-riemannian}. Using a new metric
without transporting it can change the update even
when the predictions and losses agree.

For example, let \(q=2p\) and \(\ell(p)=p^2/2\).
The unit metric in \(p\) transports to
\(\tfrac14\,dq^2\), giving gradient component \(q\).
Independently choosing the unit metric in \(q\) gives
component \(q/4\) for the same loss \(q^2/8\).
Thus the descent test for an optimizer concerns both
its prediction and its optimization data. For a
stateful optimizer the state and its transport can be
included in the updated parameter object, with the
same comparison condition on the enlarged update.
\par
\section{Two uses of quotient and local symmetry}
\label{sec:applications}

The first use chooses a nuisance action by a monoidal
topos and computes the resulting space of feasible layers.
The second uses a symmetry that varies as an internal group
and computes the admissible coefficients of a linear network.

\subsection{A central action and nuisance-invariant learning}
\label{sec:nuisance}

Let $A$ be an abelian discrete group, let $H$ be a
discrete group, and let $\chi:A\to Z(H)$ be a homomorphism.
Write $N=\chi(A)$ and $K=H/N$. The induced action of
$\mathbf B A$ on $\mathbf B H$ has geometric quotient
$q:\mathbf B H\to\mathbf B K$. Its inverse image inflates
a $K$-action along $H\twoheadrightarrow K$.
An $H$-object descends precisely when its $N$-action
is trivial \cite{topos-actions}. This is a literal action
between topoi of the kind discussed in
Section~\ref{sec:introduction}.

\begin{proposition}[Orbit features from a central quotient]
\label{prop:orbit-layer}
Let $P,W$ be $H$-sets with trivial $N$-action, and $V$ an
arbitrary $H$-set. An $H$-equivariant map $F:P\times V\to W$
corresponds uniquely to a $K$-equivariant map
$\overline F:P\times(V/N)\to W$.
For topological actions and continuous layers the same
statement holds with the orbit topology on $V/N$.
\end{proposition}

\begin{proof}
Equivariance and the hypotheses give
$F(p,nv)=nF(p,v)=F(p,v)$. Therefore $F$ is constant
on the fibers of $(p,v)\mapsto(p,[v])$ and factors.
Normality of $N$ lets $H$ act on the orbits through $K$.
The factor map is consequently equivariant for $K$,
and the converse follows by composing with the quotient.
For continuous actions, $V\to V/N$ is an open surjection
since the saturation of an open set is the union of its
translates. Its product with $P$ is also an open surjection,
so the factor map is continuous by the quotient property.
\end{proof}

Suppose now that $N$ is finite and $V,W$ are real linear
representations. For $W$ with trivial $N$-action, an
$N$-invariant \emph{linear} map out of $V$ satisfies
$L=L\Pi_N$, where
\[
 \Pi_N v=\frac{1}{|N|}\sum_{n\in N} n v .
\]
The operator is a projector onto $V^N$, and for an
$H$-representation it commutes with $H$ because $N$ is
normal. The same factorization holds for an affine map
by adding its constant part. The orbit feature $V/N$
instead retains the information needed for all invariant
nonlinear maps. To see the difference, take
$H=C_4$ acting by quarter turns on $V=\mathbb R^2$ and
$N=C_2$ acting by $v\mapsto-v$. This is obtained from
$\chi:C_2\to C_4$ with image the half turn. Here
$\Pi_N=0$. A label on the input set
$\{0,\pm e_1,\pm e_2\}$ indicating whether it is nonzero
is predicted exactly by the invariant square-activation
score $\|v\|^2$. If half of the distribution is at $0$
and half is uniform on the other four points, a predictor
fed only $\Pi_N v$ is constant. Its minimum mean square
error is $1/4$, the variance of the label. The nonlinear
invariant predictor has zero error. The quotient
determines which information may be discarded, and an
implementation still has to choose an orbit encoding.

There is a related operation on predictors rather than
input features. If $\mu$ is the joint distribution of
input and label, assume $\mu$ is invariant under
$(v,y)\mapsto(nv,y)$. For a vector score $s$ and a loss
$\ell(y,-)$ convex in the score, define
$\overline s(v)=|N|^{-1}\sum_n s(nv)$.
For measurable scores and a nonnegative loss with finite risk,
Jensen's inequality gives
\begin{align*}
 \mathbb E_\mu[\ell(y,\overline s(v))]
 &\leq \frac1{|N|}\sum_{n\in N}
             \mathbb E_\mu[\ell(y,s(nv))]\\
 &=\mathbb E_\mu[\ell(y,s(v))].
\end{align*}
The averaged predictor descends to $V/N$ by invariance.
If $s$ had an additional $H$-equivariance with $N$ acting
trivially on the output, the factor is $K$-equivariant as
in Proposition~\ref{prop:orbit-layer}.
This is an exact-invariance example of orbit averaging
in data augmentation \cite{chen-augmentation}.
The inequality compares realized predictors. An
architecture must contain the average, for example by
parallel evaluation, before it gives a comparison
within that architecture.

\subsection{A filter with a twisted coefficient}
\label{sec:twisted-filter}

Let the deployment base be $B=S^1=\mathbb R/(2\pi\mathbb Z)$.
In $\operatorname{Sh}(B)$ let $G$ be the locally constant
group sheaf with fiber $C_3=\mathbb Z/3$ and monodromy
$j\mapsto-j$. Only $0$ is fixed by inversion, so
$\Gamma(B,G)$ is the trivial group. The local groups are
nevertheless nontrivial.

Use real regular-representation feature fibers
$V_t=\mathbb R[C_3]$ with basis $(e_j)_{j\in C_3}$.
The monodromy on the bundle $V$ is the orthogonal
permutation $S e_j=e_{-j}$. Let $R e_j=e_{j+1}$
be the group shift in a local lift of the angular
coordinate. Then $SRS^{-1}=R^{-1}$, which makes the
regular action compatible with the group sheaf's
transition. The sheaf of smooth sections of $V$
is an internal $G$-object. Its fiberwise $G$-linear
maps define equivariant linear layers in the relative
classifying topos.

\begin{proposition}[The coefficient bundle of equivariant filters]
\label{prop:twisted}
The bundle of fiberwise linear intertwiners from $V$
to itself is $\mathbb R^2\oplus L$ over $S^1$, where
$L$ is the real M\"obius line and $\mathbb R^2$ denotes
a trivial rank-two bundle. A flat filter, meaning
locally constant coefficients in the given flat
frames, has two free real coefficients. Smooth
filters can also use the third coefficient as a
section of $L$.
\end{proposition}

\begin{proof}
A linear map commuting with $R$ is determined by
its value $w$ on $e_0$, since its values on the other
basis vectors must be $Rw$ and $R^2w$. Every $w$
gives a map by this rule. Hence the commutant consists
precisely of $c_0 I+c_1R+c_2R^2$. Changing frame by
$S$ exchanges $c_1,c_2$. In the basis
$I,R+R^2,R-R^2$ it is the identity on the first
two directions and multiplication by $-1$ on the
third. This gives the bundle splitting and the
flat coefficient count.
\end{proof}

The distinction affects model capacity. Using the
global section group alone would impose no commutation
conditions and would allow all nine entries of a local
matrix. Requiring internal equivariance leaves three
local coefficients. Choosing flat global coefficients
ties the two shifted terms together, leaving two.
For smooth models the third coefficient $b(t)$ can
be any smooth function on the lifted angle satisfying
$b(t+2\pi)=-b(t)$. For example
$b(t)=\cos(t/2)$ gives a well-defined global filter
$b(t)(R-R^2)$. Every continuous such $b$ has a zero:
if $b(t)$ is nonzero, the values at $t$ and
$t+2\pi$ have opposite signs. Hence a model that
requires this coefficient to stay nonzero can work over
the orientation double cover, where the pulled-back
line is trivial. Learning this coefficient in the
bundle and learning one untwisted scalar are different
parameterization problems.

\begin{figure}[tbp]
\centering
\(
\begin{tikzcd}[column sep=4.5em,row sep=2.2em]
 V_t \arrow[r,"A_t"] \arrow[d,"S"'] & V_t \arrow[d,"S"]\\
 V_{t+2\pi}\arrow[r,"A_{t+2\pi}"'] & V_{t+2\pi}
\end{tikzcd}
\)
\caption{A filter descends when it intertwines the change of
frame. Conjugation by the transition $S$ reverses the
coefficient of $R-R^2$.}
\label{fig:twisted-filter}
\end{figure}

A coordinatewise activation commutes with the local shift
and the frame permutation. Consequently one may stack
these linear filters with such activations to obtain a
global nonlinear layer. A sum of the three output
coordinates then gives a scalar readout, invariant under
both the local group and the frame transition.

A simple regression calculation gives a training
interpretation with a specified observation law.
In a local frame draw $x=e_j$ uniformly over $j$,
draw a hidden sign $\varepsilon\in\{\pm1\}$ fairly
and independently, and observe the target
$y=R^\varepsilon x$. Inverting the frame sends
$(x,y,\varepsilon)$ to $(Sx,Sy,-\varepsilon)$, so
the law descends even though a deterministic forward
shift $R$ does not define a flat global filter.
The least-squares estimator conditional on $x$ is
$(R+R^2)x/2$, an admissible flat filter.
The two conditional target vectors are distinct basis
vectors, so each lies squared distance $1/2$ from
their mean. The best expected squared error using
only $x$ is therefore $1/2$. The rule $(x,\varepsilon)\mapsto R^\varepsilon x$ has zero error
if the sign is supplied with its frame transition. The ambiguity here
is a stated property of the observation law.
Monodromy supplies the transformation of the hidden
sign and the constraint on flat filters.
\par
\section{Conclusion}

The literature provides several operational cases of learning
with symmetry-sensitive structure. Nominal automata support
active queries over infinitely many names with finitely
described orbits. Sheaf neural networks use local incidence
geometry, and groupoid-equivariant networks use transformations
that are defined only on parts of the input space. Topos and
stack semantics also supply formal descriptions of network
behavior and symbolic reasoning. The type of action differs
across these examples.

The two models in Section~\ref{sec:applications} give concrete
consequences of following the type of action through to the
parameters and observations. A central quotient identifies
the input orbit as the information available to every
nuisance-invariant predictor. The circle group sheaf imposes
a transition law on one learnable coefficient, even though
its global group of sections is trivial. An equivariant
optimizer requires compatible metric, loss, and data transport
in addition to compatible layers. The topological statements
describe admissible realizations. Prediction error and
optimization depend on the observation law and update rule
specified for the model.
\par
\begin{thebibliography}{99}
\small\raggedright

\bibitem{abramsky-contextuality}
S. Abramsky and A. Brandenburger. The sheaf-theoretic structure
of non-locality and contextuality.
\emph{New J. Phys.} 13 (2011), 113036.
\href{https://doi.org/10.1088/1367-2630/13/11/113036}{DOI}.

\bibitem{amari-natural}
S.-i. Amari. Natural gradient works efficiently in learning.
\emph{Neural Computation} 10(2) (1998), 251--276.
\href{https://doi.org/10.1162/089976698300017746}{DOI}.

\bibitem{topos-actions}
Anonymous. \emph{Actions of Toposes on Toposes:
Geometric Actions, Internal Symmetry, and Quotient Constructions}.
Companion manuscript, submitted to Higher Categorical Semantics, 2026.

\bibitem{barbe-topos-automata}
K. Barb\'e. \emph{Learning Functorial Automata in Topoi}.
Research internship report, IRIF, 2024.
\href{https://www.irif.fr/~kbarbe/files/msc.pdf}{Report}.

\bibitem{belfiore-bennequin}
J.-C. Belfiore and D. Bennequin.
Topos and stacks of deep neural networks. Preprint, version 3, 2022.
\href{https://arxiv.org/abs/2106.14587}{arXiv:2106.14587}.

\bibitem{bodnar-nsd}
C. Bodnar, F. Di Giovanni, B. P. Chamberlain, P. Li\`o, and M. M. Bronstein.
Neural sheaf diffusion: A topological perspective on heterophily
and oversmoothing in GNNs.
In \emph{Advances in Neural Information Processing Systems} 35, 2022.
\href{https://papers.neurips.cc/paper_files/paper/2022/hash/75c45fca2aa416ada062b26cc4fb7641-Abstract-Conference.html}{Proceedings}.

\bibitem{caramello-raven}
O. Caramello. \emph{Relative Toposes for Artificial General
Intelligence}, part 2. CentraleSup\'elec, July 2025.
\href{https://www.oliviacaramello.com/Talks/RelativeToposesAGI_part2.pdf}{Course slides}.

\bibitem{chen-augmentation}
S. Chen, E. Dobriban, and J. H. Lee.
A group-theoretic framework for data augmentation.
\emph{J. Mach. Learn. Res.} 21(245) (2020), 1--71.
\href{https://jmlr.org/papers/v21/20-163.html}{Journal article}.

\bibitem{cohen-geiger-weiler}
T. S. Cohen, M. Geiger, and M. Weiler.
A general theory of equivariant CNNs on homogeneous spaces.
In \emph{Advances in Neural Information Processing Systems} 32, 2019.
\href{https://proceedings.neurips.cc/paper_files/paper/2019/hash/b9cfe8b6042cf759dc4c0cccb27a6737-Abstract.html}{Proceedings}.

\bibitem{cohen-welling}
T. Cohen and M. Welling. Group equivariant convolutional networks.
In \emph{Proceedings of ICML}, PMLR 48 (2016), 2990--2999.
\href{https://proceedings.mlr.press/v48/cohenc16.html}{Proceedings}.

\bibitem{cruttwell-gradient}
G. S. H. Cruttwell, B. Gavranovi\'c, N. Ghani, P. Wilson, and F. Zanasi.
Categorical foundations of gradient-based learning.
In \emph{Programming Languages and Systems}, ESOP 2022, 1--28.
\href{https://doi.org/10.1007/978-3-030-99336-8_1}{DOI}.

\bibitem{dehaan-mesh}
P. de Haan, M. Weiler, T. Cohen, and M. Welling.
Gauge equivariant mesh CNNs: Anisotropic convolutions on geometric graphs.
In \emph{Proceedings of ICLR}, 2021.
\href{https://openreview.net/forum?id=Jnspzp-oIZE}{Proceedings}.

\bibitem{fong-backprop}
B. Fong, D. I. Spivak, and R. V. Tuy\'eras.
Backprop as functor: A compositional perspective on supervised learning.
In \emph{Proceedings of LICS}, 2019.
\href{https://hdl.handle.net/1721.1/123513}{Author manuscript}.

\bibitem{fritz-markov}
T. Fritz. A synthetic approach to Markov kernels, conditional
independence and theorems on sufficient statistics.
\emph{Adv. Math.} 370 (2020), 107239.
\href{https://doi.org/10.1016/j.aim.2020.107239}{DOI}.

\bibitem{hansen-gebhart-snn}
J. Hansen and T. Gebhart.
Sheaf neural networks. Preprint, 2020.
\href{https://arxiv.org/abs/2012.06333}{arXiv:2012.06333}.

\bibitem{hansen-ghrist-sheaves}
J. Hansen and R. Ghrist.
Toward a spectral theory of cellular sheaves.
\emph{J. Appl. Comput. Topol.} 3 (2019), 315--358.
\href{https://doi.org/10.1007/s41468-019-00038-7}{DOI}.

\bibitem{heunen-qbs}
C. Heunen, O. Kammar, S. Staton, and H. Yang.
A convenient category for higher-order probability theory.
In \emph{Proceedings of LICS}, 2017.
\href{https://doi.org/10.1109/LICS.2017.8005137}{DOI}.

\bibitem{ibort-groupoid-cnn}
A. Ibort, M. Jimenez-Vazquez, and J. M. Perez-Pardo.
Theory for groupoid equivariant neural networks: An approach
for steerable CNNs on bounded domains.
Preprint, 2026.
\href{https://arxiv.org/abs/2609.25987}{arXiv:2609.25987}.

\balance
\bibitem{iwaniack-automata-toposes}
V. Iwaniack. Automata in toposes, and general Myhill--Nerode theorems.
Preprint, 2023.
\href{https://arxiv.org/abs/2307.14855}{arXiv:2307.14855}.

\bibitem{jia-topos-survey}
Y. Jia, G. Peng, Z. Yang, and T. Chen.
Category-theoretical and topos-theoretical frameworks
in machine learning: A survey.
\emph{Axioms} 14(3) (2025), 204.
\href{https://doi.org/10.3390/axioms14030204}{DOI}.

\bibitem{lafforgue-eth-ai}
L. Lafforgue. \emph{Some Possible Roles for AI of Grothendieck Topos Theory}.
ETH Z\"urich, 2022.
\href{https://www.laurentlafforgue.org/Expose_Lafforgue_topos_AI_ETH_sept_2022.pdf}{Talk slides}.

\bibitem{lee-riemannian}
J. M. Lee. \emph{Introduction to Riemannian Manifolds}.
Second edition. Springer, 2018.
\href{https://doi.org/10.1007/978-3-319-91755-9}{DOI}.

\bibitem{maclane-moerdijk}
S. Mac Lane and I. Moerdijk.
\emph{Sheaves in Geometry and Logic: A First Introduction to Topos Theory}.
Springer, 1992.
\href{https://doi.org/10.1007/978-1-4612-0927-0}{DOI}.

\bibitem{moerman-nominal-learning}
J. Moerman, M. Sammartino, A. Silva, B. Klin, and M. Szynwelski.
Learning nominal automata.
In \emph{Proceedings of POPL}, 2017, 613--625.
\href{https://doi.org/10.1145/3009837.3009879}{DOI}.

\bibitem{simonovsky-ecc}
M. Simonovsky and N. Komodakis.
Dynamic edge-conditioned filters in convolutional neural networks on graphs.
In \emph{Proceedings of CVPR}, 2017, 3693--3702.
\href{https://openaccess.thecvf.com/content_cvpr_2017/html/Simonovsky_Dynamic_Edge-Conditioned_Filters_CVPR_2017_paper.html}{Proceedings}.

\bibitem{singer-wu-vdm}
A. Singer and H.-T. Wu.
Vector diffusion maps and the connection Laplacian.
\emph{Comm. Pure Appl. Math.} 65(8) (2012), 1067--1144.
\href{https://doi.org/10.1002/cpa.21395}{DOI}.

\bibitem{smithe-bayes}
T. St. Clere Smithe.
Bayesian updates compose optically. Preprint, 2020.
\href{https://arxiv.org/abs/2006.01631}{arXiv:2006.01631}.

\bibitem{spivak-learners-languages}
D. I. Spivak. Learners' languages.
\emph{Electron. Proc. Theor. Comput. Sci.} 372 (2022), 14--28.
\href{https://doi.org/10.4204/EPTCS.372.2}{DOI}.

\bibitem{vanoostrum-natural}
J. van Oostrum, J. M\"uller, and N. Ay.
Invariance properties of the natural gradient in overparametrised systems.
\emph{Information Geometry} 6 (2023), 51--67.
\href{https://doi.org/10.1007/s41884-022-00067-9}{DOI}.

\bibitem{villani-transformers-withdrawn}
M. J. Villani and P. McBurney.
The topos of transformer networks. Withdrawn preprint, 2024.
\href{https://arxiv.org/abs/2403.18415}{arXiv record}.

\end{thebibliography}
\end{document}
