Independent scholarship · Open accessThe monthly digest

Indexed text attachment

representation-theory-of-parametrized-neural-models-on-spaces.tex

Source for Representation Theory of Parametrized Neural Models on Spaces

View paper · Download original file

Lines 121–220 of 1343. Source text is displayed literally.

121\end{@twocolumnfalse}122]123\thispagestyle{plain}124 125\section{From statistical families to representations}126 127A supervised model commonly starts with a parameter manifold $P$, a128feature space $X$, and a map $f(p,x)$ valued in labels, predictions, or129probability distributions. A convolutional network has additional130structure. Features may live in bundles on an observation space, and a131symmetry can transport both features and weights. Two local weights can132also present the same model via a change of frame. These cases require133tracking evaluation, transport, and equivalences of presentations.134 135\emph{Representation Theory in Higher Categories} develops ordinary136and coherent group actions,137induced feature spaces, and equivariant convolutions138\cite[Chs.~17, 20, 27--28]{reprbook}. We start with its idea of encoding139actions by functors and relate evaluation, symmetry, and descent.140Evaluation records what a parameter computes. Representation theory141controls how evaluations transform. Descent determines when local142evaluations and parameter presentations assemble across a base.143Theorem~\ref{thm:relative-representation} gives the general comparison.144Theorem~\ref{thm:harmonic} computes the moving-parameter part in a145compact linear setting.146 147Parametrized composition and supervised learning have categorical148formulations through learners, lenses, and reverse differentiation149\cite{fong-spivak-tuyeras,cruttwell}. Markov categories organize channels150and sufficient statistics \cite{fritz-2020}, while equivariant151convolution on homogeneous spaces is described by induced152representations \cite{cohen-geiger-weiler}. Here we apply representation153constraints to the evaluation families themselves. This permits the same154typing of a manifold layer, a statistical decoder, and a locally presented155system of layers.156 157\section{Relative parametrized models}158\label{sec:relative}159 160Let $\E$ be an $\infty$-topos and $B$ an object of $\E$.161Possible choices are the $\infty$-category $\Spc$ of spaces and the162$\infty$-topos of sheaves of spaces on a smooth site. Work first163in $\overB$. For $X,Y$ over $B$, write $[X,Y]_B$ for their164exponential in that slice. Its evaluation map is165$[X,Y]_B\times_B X\to Y$ over $B$. A choice of site specifies which166families of sections are admitted. Smooth manifolds will be embedded167by their smooth functors of points.168 169\begin{definition}[A parametrized layer]170A layer from $X$ to $Y$ over $B$ consists of $P\to B$ and an171evaluation $f:P\times_B X\to Y$. A map to $(Q,g)$ is a morphism172$u:P\to Q$ over $B$ with a specified homotopy173$f\simeq g(u\times_B\id_X)$ over $B$. Higher morphisms are the174compatibilities in the mapping space of the resulting slice.175\end{definition}176 177Write $r_f:P\to[X,Y]_B$ for the transpose of $f$. Given a second178layer $(Q,g):Y\to Z$, the series composite has parameter179$P\times_B Q$ and evaluation $(p,q,x)\mapsto g(q,f(p,x))$.180Here and below parametric formulas describe generalized points in the181slice. A shared parameter $P$ can be imposed by reparametrization along182the diagonal $P\to P\times_B P$. More general tying is expressed by a183map $P\to\prod_{i,B}P_i$ into the product of layer parameters.184 185\begin{theorem}[Relative representation and descent]186\label{thm:relative-representation}187The $\infty$-category of layers from $X$ to $Y$ is188$\E_{/[X,Y]_B}$. On these categories, series composition is189induced by internal composition of exponentials and respects190reparametrizations. The construction commutes with pullback in $B$.191For an effective epimorphism $U\to B$ with \v{C}ech nerve $U_\bullet$,192the global layer category is equivalent to193\begin{equation}194 \label{eq:layerdescent}195 \lim_{[n]\in\boldsymbol\Delta}196   \E_{/[X_n,Y_n]_{U_n}},197 \qquad X_n=X\times_B U_n,\quad Y_n=Y\times_B U_n .198\end{equation}199\end{theorem}200 201\begin{proof}202Currying yields the equivalence of mapping spaces203\[204 \Map_B(P\times_B X,Y)\simeq\Map_B(P,[X,Y]_B).205\]206It is natural in $P$. The two definitions of a morphism of207layers therefore give the same homotopy fiber of a map of mapping208spaces, including its higher simplices. The relative double slice209$(\overB)_{/[X,Y]_B}$ is $\E_{/[X,Y]_B}$.210 211Evaluation in series transposes to a morphism212\[213 [X,Y]_B\times_B[Y,Z]_B\longrightarrow [X,Z]_B.214\]215It sends the classifying maps of the two layers to the classifying216map of their composite. The identity is classified by the transpose217of $\id_X$ with parameter $B$. The associativity and unit218compatibilities follow by uncurrying the iterated compositions.219Both ways to combine a string yield evaluation in the same order220on the product of its parameters. All product comparisons come