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 1063–1162 of 1343. Source text is displayed literally.
1063while the map from smooth weights to observable laws loses its1064entire first derivative at the origin.1065\par1066\section{Equivariant comparison and stable features}1067\label{sec:stable}1068 1069At the level of an $\infty$-category, an action of a1070discrete group $G$ on features takes the form of a functor1071from the groupoid $BG$. A layer carrying symmetry specifies1072a natural transformation from its source action to its1073target action. If its parameter and feature objects can be1074combined by a tensor product, the source is the corresponding1075diagonal action.1076 1077For a concrete stable target, take $D(k)$ for a commutative1078ring $k$. A linear layer from $X$ to $Y$ with linear parameter1079$W$ is now a map $W\otimes_k^L X\to Y$. A coherent $G$-action1080on these objects equips the ordinary mapping spectrum1081$\operatorname{map}_{D(k)}(W\otimes_k^L X,Y)$ with conjugation1082transport. In a strict chain model, an element $g$ sends1083$\varphi$ to1084$\rho_Y(g)\varphi(\rho_W(g^{-1})\otimes\rho_X(g^{-1}))$.1085The mapping spectrum of equivariant layers is1086\begin{equation}1087 \label{eq:homotopy-fixed}1088 \begin{aligned}1089 &\operatorname{map}_{\Fun(BG,D(k))}1090 (W\otimes_k^L X,Y)\\[-2pt]1091 &\qquad\simeq1092 \operatorname{map}_{D(k)}(W\otimes_k^L X,Y)^{hG}.1093 \end{aligned}1094\end{equation}1095The superscript $hG$ denotes the limit of a $BG$-action.1096The formula also holds without a parameter, by taking the1097trivial tensor unit $W=k$.1098 1099To see \eqref{eq:homotopy-fixed}, the homotopy end for natural1100transformations resolves a transformation along strings of1101group arrows. In cosimplicial degree $n$ the term is a product1102over $G^n$ of copies of the underlying mapping spectrum.1103Transporting the source of each component along its1104associated group arrows converts the two outer cofaces into1105conjugation at one end and deletion at the other.1106The inner cofaces compose successive arrows. This is the1107cosimplicial replacement for the homotopy limit on the1108right-hand side. Taking the totalization imposes all the1109coherences of an equivariant layer1110\cite[Chs.~17--18]{reprbook}.1111 1112The formula compares intertwiners between representations1113in a stable setting. It can model layers between complexes of1114features, for example when grading records features and relations1115between them. It introduces no probability law by itself.1116A stochastic decoder still requires its own evaluation and law.1117 1118\begin{example}[A split underlying comparison with an1119equivariant obstruction]1120Let $G=C_2$ and $k=\mathbb F_2$. Give $k$ the trivial action,1121and let $R=k[C_2]\simeq k[u]/(u^2)$ with $u=g-1$.1122The exact sequence of $R$-modules1123\[1124 0\longrightarrow k\xrightarrow{\,1\mapsto u\,}1125 R\xrightarrow{\,\varepsilon\,}k\longrightarrow 01126\]1127splits over $k$. An $R$-linear splitting would choose a lift1128of $1$ in $R$ killed by $u$. Every such lift is $1+au$, and1129$u(1+au)=u\ne0$. Hence the sequence gives a nonzero class of1130equivariant maps $k\to k[1]$, becoming zero when the1131action is forgotten. The free resolution obtained by repeating1132multiplication by $u$ shows1133$\Ext^j_R(k,k)\simeq k$ for every $j\ge 0$.1134This gives the positive cohomological degrees in the homotopy1135fixed point spectrum of the endomorphism spectrum of $k$.1136\end{example}1137 1138The regular representation in the example and the trivial1139module are objects of the same stable coefficient category.1140To identify the action category with derived $k[G]$-modules,1141evaluation at the unique object of $BG$ is conservative and1142preserves colimits. Its left adjoint is the free action, whose1143underlying monad is $k[G]\otimes_k^L(-)$ with multiplication1144from $G$. The same monad comes from extension and1145restriction of scalars for $k\to k[G]$. Monadicity gives the1146identification \cite[\S4.7.3]{lurie-ha}. Over a ring where1147$|G|$ is a unit, the averaging idempotent makes the trivial1148module projective and the higher Ext groups in this example1149vanish. This clarifies the relation with classical averaging.1150With real or complex Hilbert features and continuous compact1151actions, averaging and decomposition follow instead from Haar1152integration under the analytic hypotheses of1153Section~\ref{sec:harmonics}.1154\par1155\section{Losses, derivatives, and symmetry}1156 1157The classifying map of a layer can be used for optimization1158when the parameter and feature objects carry differential1159structure. Work in smooth manifolds, or use smooth charts of1160bundles over a common base. Let1161$h=f(p,x)$, $y=g(q,h)$, and let a smooth loss have output1162covector $\lambda_y=d_y\ell$. The partial reverse derivatives