Independent scholarship · Open accessThe monthly digest

Indexed text attachment

topos-actions-in-machine-learning-and-ai.tex

Source for Topos Actions in Machine Learning and Artificial Intelligence

View paper · Download original file

Lines 107–206 of 1207. Source text is displayed literally.

107kernels. The literature identified here gives little direct treatment108of learning with a monoidal Grothendieck topos acting by geometric109morphisms. We develop two applications in that setting and its110internal counterpart. A central action gives a quotient-type criterion111for nuisance-invariant layers and distinguishes quotient features from112linear averaging. A nonconstant group sheaf over the circle yields113learnable filters whose coefficient bundle has a M\"obius summand.114Descent determines which parameter choices and update rules give115consistent models across contexts. The results are structural and116are illustrated by exact learning examples.117\par\medskip118\noindent\textbf{Keywords.}\enspace119Topos actions, equivariant learning, descent, sheaf neural networks,120nominal automata, nuisance invariance.121\end{minipage}122\vspace{17pt}123\end{@twocolumnfalse}124]125\thispagestyle{plain}126 127\section{Which notion of action enters learning?}128\label{sec:introduction}129 130Equivariance in machine learning usually begins with one group,131its representations on data and features, and a trainable map132that respects those representations. Three developments motivate133a topos description. A local model may depend on a varying134context. The observations and the symmetry groups themselves135can have nontrivial transition maps. An action can also relate136different systems of observations by geometric reindexing.137Keeping those distinctions explicit lets one say which facts138are architectural and which are consequences of descent.139 140In the geometric definition of \cite{topos-actions},141the parameter $\mathcal T$ is a monoidal Grothendieck142topos. Its action on $\mathcal X$ is a geometric morphism143$a:\mathcal T\times\mathcal X\to\mathcal X$ with coherent144multiplication and unit. Its inverse image145$a^*$ pulls an observation on $\mathcal X$ back to an observation146on the product with the symmetry parameter. A second construction147is a weak action of an ordinary group by geometric148autoequivalences. It is realized by taking $\mathcal T$ to be149the sheaf topos of the discrete locale of group elements.150Finally, a group object $G$ in a fixed topos $\mathcal E$ acts151on objects of that topos. The topos $B_{\mathcal E}G$ of internal152$G$-objects classifies compatible families with those symmetries.153A locally constant sheaf of groups illustrates why its group of154global sections may be insufficient to specify a model.155 156For a finite group in sets this distinction is visible even157before learning. The parameter topos $\mathrm{Set}^{G}$ indexed158by the underlying discrete set of elements has points indexed159by $G$. The topos of $G$-sets, denoted $\mathbf B G$, classifies160$G$-torsors. Its points form a groupoid with one isomorphism161class and automorphism group $G$. Group laws on the set of162elements induce geometric operations on the parameter topos.163A Cartesian monoidal law on $\mathbf B G$ requires $G$ to be164abelian, as proved in \cite{topos-actions}. Below we use an165actual action of $\mathbf B A$ for abelian $A$ and compute166what it says about learning a nuisance-invariant readout.167 168\subsection{External symmetry and spaces of layers}169 170Let an ordinary group $H$ act by geometric autoequivalences on171a Grothendieck topos $\mathcal E$. Write $T_h$ for the inverse172image indexed by the inverse element, so that173$T_hT_k\simeq T_{hk}$. An equivariant object $X$ has structure174$\theta_{h,X}:T_h X\simeq X$. The equivariant topos175$\mathcal E^H$ is the geometric quotient. For equivariant176objects $X,Y$, the underlying exponential $Y^X$ acquires177the conjugation action. The equivalence178$T_h(Y^X)\simeq(T_hY)^{T_hX}$ is followed by precomposition179with $\theta_{h,X}^{-1}$ and postcomposition with $\theta_{h,Y}$.180The two coherence laws follow by composing these maps on181either side of an evaluation. Consequently a layer182$P\times X\to Y$ is equivariant precisely when its transpose183$P\to Y^X$ is equivariant. The usual equation184$F(hp,hx)=hF(p,x)$ is its case in sets. It is an equation185about the family. At a fixed parameter the equation applies186to its stabilizer.187 188This elementary result makes the applications testable.189An architecture specifies a parameter object, a realization190map, and which symmetries it respects. A loss and a learning191rule require their own compatibility with those data.192Sections~\ref{sec:direct}--\ref{sec:categorical-ml} review193relevant work. The later sections develop descent for layer194evaluation and compute two specializations where the195symmetry changes what can be learned or represented.196\par197\section{Topos-theoretic precedents in learning}198\label{sec:direct}199 200The closest precedents use three different levels of structure.201Some place the states and maps of a learner in a topos of equivariant202objects. Others use a classifying topos to organize possible203interpretations and changes of context. A literal action in the204sense considered here gives a monoidal Grothendieck205topos \(\mathcal T\) a geometric action206\(\mathcal T\times\mathcal X\to\mathcal X\) on another topos. One