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 281–380 of 1207. Source text is displayed literally.

281Lafforgue \cite[slides 20--25]{lafforgue-eth-ai} proposes282organizing a data language by a geometric theory \(T\).283A map from a test topos to its classifying topos classifies a284\(T\)-model in that test topos. Taking the test topos to vary285therefore models families of interpretations and their286change of base. His proposed networks of topoi attach287geometric object categories and classifiers to different288processing stages. A transformation natural in whole289families is a candidate for a geometric morphism between290classifiers. This provides a way to formulate the transfer of291structured representations. The talk gives an architecture292proposal rather than an optimization or accuracy result.293Caramello's course on relative topoi \cite{caramello-raven}294makes the symbolic application more concrete. It writes295invariance constraints for Raven matrices in a geometric296language, including an example where an arithmetic rule on297positions is expressed by a formula indexed by the graph of298that rule. Restricted vocabularies and specified relations299make candidate interpretations comparable by proof. This is300a formal model of a class of reasoning problems and a distinct301application from training an equivariant numerical predictor.302 303\subsection{An algorithmic precedent via nominal sets}304 305Active automata learning gives an implemented example of306learning in a symmetry-sensitive setting. The nominal topos307is the category of discrete sets with continuous action of the308permutation group of countably many names, equipped with309the topology of pointwise convergence. Continuity of an orbit310map means that an element has a finite support: its stabilizer311contains the subgroup fixing some finite set of names312pointwise. The set of name pairs, for example, has two orbits,313equal and unequal pairs. Such orbit-finite structure permits314a finite orbit description even for machines with infinitely315many individually named inputs. Iwaniack316\cite{iwaniack-automata-toposes} uses the nominal topos as317an instance of topos-enriched automata and derives318Myhill--Nerode results there.319 320Moerman and collaborators \cite{moerman-nominal-learning}321adapt Angluin's query algorithm to nominal automata, including322preliminary experiments with a Haskell implementation.323The teacher answers queries about acceptance and proposes324counterexamples to incorrect hypotheses. Equivariance lets325sets of examples and hypotheses be handled by their orbit326structure. Barb\'e \cite[Thm.~4.26]{barbe-topos-automata}327generalizes this formulation in a research report.328Over an atomic topos with countable coproducts, an329alphabet and target state object each having finitely330many atom components, and an appropriate teacher,331\(\mathrm{TopL}^{*}\) terminates with a minimal automaton.332This conclusion concerns exact learning in the stated333query model. An effective implementation still depends on334the ability to manipulate the orbit or atom presentations335and to supply the teacher. These are meaningful operational336assumptions when trying to infer a structured protocol model337whose states depend on relationships among names.338\par339\section{Sheaf learning and local equivariance}340\label{sec:sheaves}341 342Two established lines of work supply concrete models for343applying topos-action ideas. Spectral sheaf learning builds344features that satisfy incidence constraints. Equivariant345network design builds maps that commute with transports.346Their intersection can be described exactly for invertible347transports on a graph.348 349\subsection{From sheaf Laplacians to learned incidence maps}350 351Hansen and Ghrist develop the spectral theory of cellular352sheaves, including Hodge Laplacians and their relation to353sheaf cohomology \cite{hansen-ghrist-sheaves}. Hansen and354Gebhart use the sheaf Laplacian in graph neural layers355\cite{hansen-gebhart-snn}. Bodnar et al.\ learn the356incidence maps from data, study how sheaf geometry changes357diffusion and oversmoothing, and compare their models on358graphs where adjacent vertices can have different labels \cite{bodnar-nsd}. An earlier direction359in graph learning uses edge-conditioned filters360\cite{simonovsky-ecc}. Edge attributes there determine361message weights. In a sheaf layer, maps from the two362endpoints into a common stalk determine which local363features agree and how disagreement contributes to the364diffusion.365 366For a finite loopless graph $\Gamma$, write $V_v$ for a vertex367stalk and $V_e$ for an edge stalk, with maps368$r_{ev}:V_v\to V_e$ at each incidence. Give the finite369dimensional stalks inner products and orient each edge370$e:u\to v$ for purposes of signs. Set371\[372  (\delta x)_e=r_{ev}x_v-r_{eu}x_u,\qquad373  L=\delta^*\delta .374\]375The Laplacian satisfies376$\langle x,Lx\rangle=\|\delta x\|^2$.377Thus $\ker L=\ker\delta$ is the vector space of global378sections, and $e^{-tL}$ converges to orthogonal projection379onto that space as $t\to\infty$. This last statement is380for the fixed unnormalized operator on a finite graph.